simultaneous functional magnetic resonance …mediatum.ub.tum.de/doc/1070878/932582.pdfsimultaneous...

TECHNICAL UNIVERSITY OF MUNICH &

MAX PLANCK INSTITUTE OF PSYCHIATRY

Simultaneous Functional Magnetic Resonance

Imaging and Electrophysiological Recordings:

Practical Application And

Methodological Approach

Dissertation

by

Kátia C. Andrade

[2011]

[This work was financially supported by the Capes Foundation, Ministry of Education of

Brazil, Caixa Postal 365, Brasília – DF 70359-970, Brazil]

TECHNISCHE UNIVERSITÄT MÜNCHEN

Physik Department E18

Simultaneous Functional Magnetic Resonance Imaging

And Electrophysiological Recordings:

Practical Application And Methodological Approach

Katia Cristine Andrade

Vollständiger Abdruck der von der

Fakultät für Physik der Technischen Universität München

zur Erlangung des akademischen Grades eines

Doktors der Naturwissenschaften (Dr. rer. nat.)

genehmigten Dissertation.

Vorsitzender: Univ.-Prof. Dr. J. Leo van Hemmen

Prüfer der Dissertation:

1. Univ.-Prof. Dr. Stephan Paul

2. apl. Prof. Dr. Sibylle Ziegler

Die Dissertation wurde am 02.03.2011 bei der Technischen Universität München eingreicht und

durch die Fakultät für Physik am 21.03.2011 angenommen.

“So much sorrow and pain

Still I will not live in vain

Like good questions never asked

Is wisdom wasted on the past […]”

(Ben Harper-“Blessed to be a witness”)

CONTENTS

FIGURES .................................................................................................................... 6

TABLES ...................................................................................................................... 7

LIST OF ABBREVIATIONS ........................................................................................ 8

ABSTRACT .............................................................................................................. 11

INTRODUCTION ...................................................................................................... 13

CHAPTER 1 ............................................................................................................. 15

Magnetic resonance imaging ............................................................................................15

Functional magnetic resonance imaging (fMRI) ................................................................21

FMRI preprocessing .......................................................................................... 24

Slice time correction ....................................................................................... 24

Motion correction ........................................................................................... 24

Normalization ................................................................................................. 25

Spatial smoothing .......................................................................................... 27

Temporal filtering ........................................................................................... 28

Statistical analysis ............................................................................................. 28

Model free analysis............................................................................................ 30

Electroencephalography (EEG) ........................................................................................32

EEG acquisition and preprocessing................................................................... 34

Simultaneous EEG and fMRI recordings ..........................................................................38

CHAPTER 2 ............................................................................................................. 44

Sleep spindles and hippocampal functional connectivity in human NREM sleep .... 44

Introduction .......................................................................................................................44

Sleep ................................................................................................................. 44

Memory and Hippocampus ................................................................................ 47

Resting-state ..................................................................................................... 49

Objective and hypotheses ................................................................................. 51

Methods ...........................................................................................................................52

Subjects ............................................................................................................. 52

FMRI and EEG acquisition ................................................................................ 52

Sleep stage rating.............................................................................................. 53

FMRI analysis .................................................................................................... 54

Sleep spindle analysis ....................................................................................... 57

Psychophysiological interaction analysis ........................................................... 58

Results .............................................................................................................................59

Sleep stage and HF functional connectivity ....................................................... 59

Subregions analysis .......................................................................................... 62

Sleep spindles ................................................................................................... 64

Discussion ........................................................................................................................66

HF connectivity to the DMN in wakefulness and sleep ...................................... 66

HF connectivity to neocortex generally higher in S2 than in SW ....................... 67

Increased HF connectivity in S2 interacts with sleep spindles ........................... 69

Limitations ......................................................................................................... 70

Conclusion ......................................................................................................... 70

CHAPTER 3 ............................................................................................................. 71

Recurrence quantification analysis allows for single-trial stimulus differentiation in evoked potential studies ........................................................................................... 71

Introduction .......................................................................................................................71

Objective ........................................................................................................... 74

Approach ........................................................................................................... 74

Methods ...........................................................................................................................75

Subjects ............................................................................................................. 75

Experimental task .............................................................................................. 75

MRI data acquisition .......................................................................................... 76

EEG acquisition and preprocessing................................................................... 77

Recurrence Quantification Analysis ................................................................... 78

Definition of ERPs and RQA time windows ....................................................... 82

Statistical analysis ............................................................................................. 83

Results .............................................................................................................................87

Explorative analysis of ERP amplitude and RQA variables ............................... 87

Cross-methods comparison of differentiation between rare and frequent tones (EEG recordings outside the scanner) ............................................................... 91

Cross-methods comparison of differentiation between subtypes of frequent tones (EEG recordings outside the scanner) ............................................................... 92

Cross-methods comparison of differentiation between rare and frequent tones (EEG recordings inside the scanner) ................................................................. 95

Cross-methods comparison of differentiation between subtypes of frequent tones (EEG recordings inside the scanner) ................................................................. 96

Correlation between EEG amplitude based measures and AvgLsum ................. 97

Discussion ........................................................................................................................99

Previous attempts of ERP single trial characterization ...................................... 99

Comparison of RQA and AA based on pooled trial responses ........................ 101

Comparison of RQA and AA based on comparisons of subsequent trials ....... 102

Influence of embedding parameters and other parameters on RQA performance ........................................................................................................................ 104

Correlation between EEG amplitude based measures and AvgLsum ............... 105

Conclusions ..................................................................................................... 106

CONCLUSION AND OUTLOOK ............................................................................. 108

APPENDIX A2 ........................................................................................................ 112

APPENDIX A3 ........................................................................................................ 124

REFERENCES ....................................................................................................... 125

ACKNOWLEDGEMENTS ....................................................................................... 148

CONTRIBUTIONS .................................................................................................. 150

FIGURES

Figure 1.1 Magnetic and magnetization vectors representation. .............................. 18

Figure 1.2. Hemodynamic response function. .......................................................... 22

Figure 1.3. Example of a auditory block-design and its activation in the auditory cortex in response to the task ................................................................................... 23

Figure 1.4. fMRI motion correction matrices. ............................................................ 25

Figure 1.5. Visual representation of linear and nonlinear transformations during the fMRI normalization step ............................................................................................ 26

Figure 1.6. Brodmann atlas and common associated function. ................................ 27

Figure 1.7. Schematic representation of the general linear model (GLM) ................ 29

Figure 1.8 Afferent inputs into the apical dendrites and cell body. ............................ 33

Figure 1.9 Electrodes positions of the scalp EEG at standardized locations ............ 35

Figure 1.10 Example of gradient artifact in a concurrent EEG recording.. ................ 40

Figure 1.11 Example of BCG artifact in a concurrent EEG recording. ...................... 42

Figure 2.1. Relationship between human sleep, level of consciousness and EEG patterns. .................................................................................................................... 46

Figure 2.2. Cytoarchitecture of the mesial temporal lobe at the level of the body of the hippocampus ............................................................................................................ 49

Figure 2.3. Examples of different resting-state networks .......................................... 50

Figure 2.4. Probabilistic cytoarchitectonic maps of the hippocampal subregions and subiculum.. ............................................................................................................... 56

Figure 2.5. Extension of the HF network during wakefulness.. ................................. 60

Figure 2.6. Main effect of sleep results of the full-factorial design employing factors sleep stage and subregion. ....................................................................................... 60

Figure 2.7. Comparison of HF functional connectivity across sleep stages .............. 62

Figure 2.8. Main effects of subregional results of the full-factorial design employing factor sleep stage and subregion .............................................................................. 62

Figure 2.9. Comparison of subregional HF functional connectivity within sleep stages.. ..................................................................................................................... 63

Figure 2.10. Activity related to fast sleep spindles in S2 ........................................... 64

Figure 2.11. PPI analysis for the interaction spindles x SUB. ................................... 65

Figure 3.1. Active auditory oddball experiment. ........................................................ 72

Figure 3. 2. Time series and recurrence plot of a random and a sinusoid system. ... 79

Figure 3.3. RQA and AA pipeline.. ............................................................................ 84

Figure 3.4 Differentiation power of AA and RQA group analysis for undistorted EEG recordings and correlation between RQA variables .................................................. 88

Figure 3.5. Topography of ERP grand averages ...................................................... 89

Figure 3.6. RQA measure AvgL using different embedding parameters m and d .. 90

Figure 3.7. Single trial correlation between AvgLsum and AA measures.................... 98

TABLES

3.1. RQA variables ................................................................................................... 81

3.2. Effect sizes for AvgL and P300 gained from undistorted EEG recordings ......... 92

3.3. Comparison of trialwise difference values (undistorted EEG recordings) .......... 93

3.4. Effect sizes for AvgL and P300 derived from EEG recordings during an fMRI scan .......................................................................................................................... 95

3.5. Comparison of trialwise difference values (EEG recordings during an fMRI scan) ................................................................................................................................. 96

A2.1. HF functional connectivity during wakefulness .............................................. 112

A2.2. Sleep stage specific functional HF connectivity changes .............................. 113

A2.3. Subregional HF functional connectivity per sleep stage. ............................... 116

A2.4 Activity related to fast sleep spindles (14 - 16 Hz) .......................................... 120

A2.5. Activity related to the interaction effect of Spindle × SUB ............................. 121

A2.6. Activity related to the interaction effect of Spindle × FD ................................ 122

A2.7. Activity related to the interaction effect of Spindle × CA ................................ 123

A3.1 Higher effect sizes found for each subject with optimized m and d values ... 124

8

LIST OF ABBREVIATIONS

AA Amplitude analysis

ACC Anterior cingulate cortex

ANOVA Analysis of variance

AvgL Average diagonal line

BCG Ballistocardiogram artifact

BOLD Blood oxygenation level dependent

CA Cornu ammonis

CBF Cerebral blood flow

CBV Cerebral blood volume

CMRO2 Cerebral metabolic rate of oxygen

CSF Cerebrospinal fluid

DET Determinism

DG Dentate gyrus

DMN “Default mode” network

EC Entorhinal cortex

ECG Electrocardiogram

EEG Electroencephalography

emf Electromotive force

EMG Electromyogram

EOG Electrooculogram

EPI Echoplanar images

EPSP Excitatory postsynaptic potential

ERP Evoked related potential

List of abbreviations 9

FFT Fast-Fourier transform

fMRI Functional magnetic resonance imaging

FOV Field of view

FWHM Full width half maximum

GLM General linear model

HF Hippocampal formation

HRF Hemodynamic response function

ICA Independent component analysis

iICA Iterative independent component analysis

IPL Inferior parietal lobule

ISI Inter-stimulus interval

KC K-complex

LAM Laminarity

MaxL Maximal diagonal line

MaxV Maximal vertical line

MEG Magnetoencephalography

MNI Montreal neurological institute

mPFC Medial prefrontal cortex

MR Magnetic resonance

NMR Nuclear magnetic resonance

NREM Non rapid eye movement

PCC Posterior cingulate cortex medial

PET Positron emission tomography

PPI Psychophysiological interactions

REM Rapid eye movement

List of abbreviations 10

rf Radiofrequency

rms Root mean square

ROI Region of interest

RP Recurrence plot

RQA Recurrence quantification analysis

RSNs Resting-state networks

RspC Retrosplenial cortex

S1 Sleep stage 1

S2 Sleep stage 2

SMA Supplementary motor area

SPECT Single-photon emission computed tomography

SPM Statistical parametric mapping

SUB Subiculum

SWS Slow wave sleep

TE Echo time

TR Repetition time

TT Trapping time

W Wakefulness

WD Wavelet denoising

11

ABSTRACT

The present work investigated two distinct analytical approaches to

simultaneous EEG/fMRI data. In the first part, functional connectivity analysis was

applied to simultaneous EEG/fMRI data to study cerebral networks and

hemodynamic correlates of sleep specific EEG features as well as their functional

significance during human sleep. In the second part, methodological aspects were

addressed, evaluating the potential of a non-linear analysis method called recurrence

quantification analysis (RQA) to extract rapidly changing dynamic features of

electrophysiological data, which may be used to improve fMRI models of brain

behaviour.

According to the standard model of memory consolidation, memory formation

initially occurs in the hippocampus and the medial temporal lobe, where new

experiences are temporarily stored. Long-term memory storage relies on off-line

transfer to and consolidation in the neocortex. This information transfer between the

hippocampus and the neocortex is believed to benefit from sleep. The goal of the first

study was to determine spontaneous functional connectivity maps of subregions of

the hippocampal formation (HF) with the rest of the brain in humans, and to evaluate

their alteration throughout non rapid eye movement (NREM) sleep. We provided first

evidence of changes in human hippocampal connectivity patterns from wakefulness

to NREM sleep. The strongest functional connectivity between the HF, especially the

subiculum output region, and neocortex was observed in sleep stage 2, while

weakest connectivity was found in slow wave sleep. Increased connectivity between

HF and neocortical regions in sleep stage 2 further appears associated to the

presence of fast sleep spindle activity in concurrent electrophysiological recordings.

This suggests an increased capacity for possible global information transfer, while

Abstract 12

reduced long-range connectivity in slow-wave sleep may reflect a functional system

optimal for segregated information reprocessing. Our data may be relevant to

differentiating sleep stage specific contributions to neural plasticity as proposed in

sleep dependent memory consolidation.

Simultaneous EEG/fMRI recording allow to combine the high EEG temporal

resolution with the unmatched fMRI spatial resolution and the flexibility of fMRI to

explore brain behaviour in various sensoric, cognitive and emotional paradigms.

Today, analysis of event related potentials (ERP) is usually performed by an

averaging of multiple EEG trials, in order to increase the signal-to-noise ratio. Thus,

the temporal resolution of the EEG is partially lost, and averaging does not consider

response changes across trials. This variability can, however, represent changes in

subject’s performance or fluctuations in attention, arousal, habituation or other

cognitive features. In the second part of this thesis, we evaluated the potential of a

nonlinear signal analysis method, recurrence quantification analysis (RQA), to

improve the characterization of single trial EEG responses. This analysis showed that

RQA is not significantly superior to conventional amplitude analysis. Despite equal

discrimination power, RQA measures were only weakly correlated with ERP

amplitudes, suggesting that additional information on single trials may be extracted

by RQA. Therefore, RQA may be used as an additional tool to obtain new insights

about the neurophysiology of the brain.

13

INTRODUCTION

Nowadays, the major techniques used to study functional brain activity are

positron emission tomography (PET), single-photon emission computed tomography

(SPECT), functional magnetic resonance imaging (fMRI), electroencephalography

(EEG) and magnetoencephalography (MEG). EEG and MEG, different from the

others, are based on electrophysiological instead of hemodynamic properties. FMRI

has the advantage over PET and SPECT of a better spatial resolution in addition to

the absence of potential risks associated with radioactive tracers. This allows

repeated measures of brain activity across various task manipulations [1]. The fMRI

temporal resolution, however, is still in the order of seconds while neuronal

processes have a millisecond scale. The indirect relationship between fMRI and the

neuronal electrical activity also raises the challenge of interpreting its findings [2;3].

EEG and MEG, on the other hand, have a high temporal resolution but lack spatial

specificity. To overcome these limitations, many research groups attempt to combine

techniques to complement each other. This is especially true for the combined use of

hemodynamically based (fMRI) and electrophysiological (EEG) techniques [4]. Since

physiological variables as vigilance, attention and anxiety cannot be fully reproduced

in two independent sessions [5;6], the simultaneous use of EEG and fMRI has

caught special attention in recent years [5;7-11].

The present work focus in two different lines of approaches to simultaneous

EEG/fMRI data; therefore, the manuscript is divided in three chapters: The first

chapter describes the technical basis and general principles of fMRI and EEG

recordings and the challenges to combine these techniques. The second and third

chapters describe the goals, background, specific methods, results and conclusion of

Introduction 14

each of the two lines of approach. Shortly, in the first project, current analyses

methods are applied to simultaneous EEG/fMRI data in order to study functional

brain connectivity and hemodynamic correlates during human sleep. In the second

study; more methodological aspects are addressed, evaluating the potential of a non-

linear method analysis called recurrence quantification analysis (RQA) to extract

rapidly changing dynamic features of electrophysiological data which may be used to

better study fMRI models.

15

CHAPTER 1

Magnetic resonance imaging

In 1946, Bloch and Purcell independently described for the first time the

physical phenomenon of nuclear magnetic resonance (NMR), where they measured

the precession effect of the spin around a magnetic field [12;13]. This achievement

leaded both to receive the Nobel Prize in physics in 1952. In research, NMR was a

rapidly growing field in organic chemistry and biochemistry. However, the idea to use

NMR to obtain spatial information of an object came just in 1973 from Lauterbur [14].

The NMR image derives from radio waves signals coming from the proton in

hydrogen nuclei as a result of its interaction with external magnetic fields. The

contrast in such an image is provided by differences in the signal intensity from these

nuclei in various tissues, and is determined by several factors like proton density or

relaxation times T1, T2 and T2*.

When protons are placed in a static magnetic field 0B , their nuclear spins will

just have two possible alignments, parallel or anti-parallel to the field. Protons aligned

in parallel are at lower energy level then the anti-parallels. As it is common in nature

to seek for lower energy levels, a slightly exceeding amount of spins will be placed in

the parallel direction. At body temperature, when the difference of energy between

the states E∆ is much lower than the thermal energy, the fraction of parallel and anti-

parallel spins is given by the Boltzmann distribution [15]:

Spin excess kT

EN

2

∆≈ , (1.1)

Chapter 1 – Magnetic resonance imaging 16

N is the total number of spins present in the sample, T is the absolute temperature

and k the Boltzmann’s constant. In NMR, E∆ is found to be proportional to the

strength of the magnetic field:

0BE γh=∆ , (1.2)

h is the Planck’s constant h divided by π2 and γ the gyromagnetic constant of the

respective nucleus.

Due to the larger number of parallel spins, a net magnetization vector is

created )(Mr

, and it is this vector that, through some experimental manipulation, give

rise to an induced current in a receiver coil, required to generate a NMR image [15].

From Faraday’s law of induction, in order to induce a current in a receiver

coil, it is necessary that the magnetization vector (the magnetic moment vector

density due the spin population) generates a variable magnetic field nearby the coil.

In the thermal equilibrium, Mr

is constant and aligned parallel to the magnetic field.

To disturb this equilibrium, energy is required to provoke transitions between the

energy levels. This is achieved by irradiating the protons with radiofrequency (rf)

magnetic fields. However, according to quantum theory, the transition between

energy levels just happens if the amount of energy exactly matches the difference of

energy between the states, in this case, E∆ . This means that the photons of the rf

field should have a frequency that satisfies:

νhE =∆ , (1.3)

h is the Planck’s constant and ν is the radio wave frequency. This leads to:

0000

22B

Bγω

πω

πγ

ν =⇒== , (1.4)


0ω is the angular frequency of the photon, referred as “Larmor frequency”. Therefore,

the resonance phenomenon is the transition of protons between the energy levels by

applying a rf with angular frequency 0ω .

A classical way to describe NMR is imagining the proton as a tiny spinning

charge. According to classical physics, with protons in the presence of a static

magnetic field )(Br

, the magnetic field will exerted a torque )(τr

that twists the

proton’s magnetic moment into a parallel alignment with the field. Because of the

spin, the torque will not twist the magnetic moment vector into alignment with the

magnetic field but will induce precession around the magnetic field direction while

maintain a constant angle.

The relationship between the magnetic moment )(µr

and the spin angular

moment )(Lr

is:

Lrr

γµ = . (1.5)

For a volume element V small enough that the external fields are

approximately constant over V but still contain a large number of protons, the

magnetic moment vector density due to the spin population Mr

will be:

∑=i

iV

M µrr 1

, (1.6)

With Ni ...1= (total of spins in V ). If no interaction of the protons with their

environment is assumed, the equation of motion can be simply written as:

( )BMdt

Md rrr

v×== γτ . (1.7)

Now, suppose the configuration in Figure 1.1 where Br

is a constant with

module 0B at z direction and Mr

with module 0M written as:


[ ]zyxMM ˆ)cos(ˆ)sin()sin(ˆ)cos()sin(0 θϕθϕθ ++=r

. (1.8)

Figure 1.1 Magnetic and magnetization vectors representation. Vectors Br and M

r in the

orthogonal )ˆ,ˆ,ˆ( zyx plane.

From (1.7):

.0

),cos()sin()cos()sin()sin()cos(

),sin()sin()sin()sin()cos()cos(

0

0

=

−=+

=−

dt

d

Bdt

d

dt

d

Bdt

d

dt

d

θ

ϕθγϕ

ϕθθ

ϕθ

ϕθγϕ

ϕθθ

ϕθ

(1.9)

The solution for (1.9) is:

.0

0

=

−=

dt

d

Bdt

d

θ

γϕ

. (1.10)

Note that 0ωωϕ

=≡dt

d, where ω is the angular frequency M

r rotates around z . The

minus in (1.10) means that Mr

rotates clockwise, with time, in the negative ϕ

direction. Thus, when a constant magnetic field is applied, the magnetization will

precess with the Larmor frequency and in a constant angle θ related to Br

.


To produce a current in a coil, a magnetic moment needs to generate a

variable magnetic field nearby the coil, thus 0≠dt

dθ. To accomplish this, the

magnetization can be rotated away from its alignment along the Br

axis by applying a

radiofrequency magnetic field for a short period of time (rf pulse) with the Larmor

frequency. After the magnetization has been rotated into the transverse plane, it will

tend to relax to equilibrium along the direction of the static field Br

. The NMR signal

corresponds to the voltage induced in a receiver coil from time-varying magnetic flux

which is produced by this rotating magnetization.

The goal of imaging is to correlate a series of signal measurements with

spatial locations of the various sources. Lauterbur proposed in 1973 [14] to use

magnetic field gradients (spatially changing magnetic fields) to label the signal from

protons according to their spatial position. If the magnetic field is dependent, for

example, on its x coordinate, so is the Larmor frequency, thus it is possible to

selectively excite protons belonging to a plan perpendicular to z by choosing an

excitation pulse having the exact frequency corresponding to the Larmor frequency of

the chosen plane (slice selection). During data acquisition, a magnetic field along the

x -axis results in a precession frequency which depends on the precise localization of

the respective spins. Therefore, the induced current in the receiver coil comprises a

sum of different precession frequencies, which can be decomposed by a fast Fourier

transformation. The x gradient is also called the read-out gradient. The y coordinate

of the protons is labeled with a gradient pulse along the y direction (frequency

encoding) before the acquisition of the signal. This gradient makes protons situated

in different y positions present different phases. By repeating this frequency


encoding step with successively increasing y gradient strength, the signal obtained

during the frequency encoding can be decomposed later on, again using Fourier

transform.

So far, the interaction between spins and their environment was neglected;

however, it is necessary to consider both internal magnetic and electric fields. These

fields influence the magnetization, a process known as relaxation. A set of equations

was proposed by Bloch to describe the temporal evolution of the magnetization that

incorporates the relaxation and precession effects [12]. The relaxation back to the

longitudinal orientation is characterized by a time constant T1 and the transverse loss

of phase coherence, perpendicular to the field 0B , by T2. The transverse relaxation

T2, based on spin-spin interactions, is of special importance for fMRI. In a perfectly

homogeneous magnetic field, the transverse relaxation would decay exponentially

with a time constant of T2. Local field inhomogeneities, however, disturb the phase

coherence and increase the speed of this transversal relaxation process, resulting in

a decay of constant T2* shorter than T2. Such field inhomogeneities may be caused

by external factors like an inhomogeneous magnetic field or by anatomical features,

like e.g. air filled cavities close to brain tissue causing transitions in magnetic

susceptibility. In the brain, these inhomogeneities may also depend on altered

physiological states, as the local blood supply and the blood oxygenation are altered

by neural activity [3]. Therefore, macroscopic tissue properties can be derived from

T2*.

Chapter 1 21

Functional magnetic resonance imaging (fMRI)

In 1990, Ogawa et al. [16;17] observed for the first time that the a specific

NMR contrast is sensitive to the oxygenation level of the blood, then termed “blood

oxygenation level dependent” (BOLD) contrast. Since the hemoglobin is a

diamagnetic molecule when oxygenated (oxyhemoglobin) but paramagnetic when

deoxygenated (deoxyhemoglobin), the presence of deoxyhemoglobin affects the

local magnetic susceptibility by creating small distortions in the magnetic field and

consequently, decreasing T2*. Upon neural activation, the increases in cerebral blood

flow (CBF) are about ~ 2-4 times larger than increases in cerebral metabolic rate of

oxygen (CMRO2). As a result, the supply of oxygen to the capillary bed is

substantially increased, but only a fraction of oxygen is removed from the blood to

sustain metabolism, so in total, the blood is locally more oxygenated and T2* is

increased [2]. This makes it possible to define the anatomical locus of the activation

by means of fMRI. However, it is important to be aware that the BOLD signal does

not directly measure neural activation but rather reflects a hemodynamical cascade

of changes in CBF, CMRO2 and cerebral blood volume (CBV). This neurovascular

coupling produces a complex MR (magnetic resonance) signal function related to the

neural response to stimuli: the hemodynamic response function (HRF) which can be

used to analyze amplitude changes in the MRI images in the order of 0.5 – 3 %, see

Figure 1.2.

FMRI studies are mostly analyzing “evoked activity”, in other words, neuronal

response to brief stimuli or activity during a task performance in association to such

stimuli. Several sophisticated statistical methods have been developed to identify

significant signal changes which are taken as representation for changes in neural

activity at that location. The first step of such an experiment is to carefully plan the

Chapter 1 – Functional magnetic resonance imaging 22

experimental paradigm to be worked on by the subject. The paradigm corresponds

to the series of tasks to be performed during the fMRI measurement.

Figure 1.2. Hemodynamic response function following a hypothetical short duration stimulus

(represented by the red bar). The maximum amplitude of the MR signal changes is in the order

of 0.5-3% and has a delay of ~ 5 s after stimulus presentation. Graph created using SPM

software version 5 (www.fil.ion.ucl.ac.uk/spm).

Currently, fMRI typically involves acquisition of a continuous series of scans

with BOLD contrast images of the whole brain collected every 1 – 2 s (depending on

the time of acquisition or repetition time – TR). A scan consists of several thousand

data points, each of which is derived from a small cube of brain tissue or voxel

(volume element). Due to the low amplitude of such evoked signals (Figure1.3),

usually the subject has to perform the same task several times to increase the

statistical power in detecting these signals. A typical example is a block-design,

where the subject maintains cognitive engagement in a task altering with periods

when a different condition is required. Figure 1.3a represents a block-design, with

the subject being exposed to auditory stimulations during 30 s intercalating with silent

periods of 30 s. Figure 1.3b shows brain regions in the auditory cortex in which the

BOLD signal fits to the hypothesis of an increase during the auditory stimulation and

a return to the baseline during the silent periods (red line in Figure 1.3a). Depending

on the research question, the paradigm can be much more complex, with a design in

an event-related (short periods of task intercalated with long control periods, the HRF


of a single event can be detected and analyzed in detail) or mixed fashion (a

combination of events closely presented, intermingled with control condition), or a

behaviorally driven one (for example, resting-state studies, to be better explained

later).

Figure 1.3. (a) Example of a block-design with 4 blocks of auditory stimulation during 30 s

intercalating with 5 silent blocks of 30 s. The red line represents the expected HRF. (b)

Activation in the auditory cortex in response to the task in (a) given as red cluster overlayed on

an anatomical MR image. Color bar represents t-values depiction statistical significance of the

derived action clusters.

Intuitively, after recording fMRI data, the first attempt would be to statistically

compare the temporal signal of each voxel in the image with the expected HRF.

However, during recording, the data can be corrupted by movement of the subject,

heartbeat, breathing, inhomogeneities of the magnetic static field, differences in the

acquisition time of the images and other factors. Therefore, before reaching a final

result image as displayed in Figure 1.3b, many preprocessing steps are needed to

eliminate artifacts and signal changes not associated with the experimental design

and to maximize the sensitivity of the statistical analysis. The basic steps include

slice time correction, motion correction, spatial and temporal filtering. For a group

analysis a normalization process is also needed.


FMRI preprocessing

Slice time correction

The scan of a whole brain volume is not instantaneously measured. Instead,

several slices are recorded two-dimensionally and then reconstructed as a three-

dimensional image. Hence, each slice is recorded at a slightly different time with

respect to a given stimulus which should be corrected for. One could, for example,

assume different HRFs for each slices, however, the most common procedure is to

interpolate the time series to “align” the slices prior to analysis. The advantage of

interpolation is that inferences concerning differences in HRF amplitude between

slices are less likely to be confounded by different slice acquisition times. However,

this approach is limited by the Nyquist frequency ( TR21 ) which makes the

interpolation unable to contain experimental power at frequencies above this

frequency [18].

Motion correction

For fMRI data analysis it is also necessary that each voxel is spatially

invariant during the whole experiment. Variation of intensity in the same voxel

between repeated volume acquisitions is primarily supposed due to changes in

cerebral physiology. However, during the experiment, movement of the head by the

subject, physiological motions related to respiration and pulsation of the blood can

result in shifting the signal time-courses of neighboring voxels. Thus, when preparing

the experiment, the comfort of the subject and demand of the task should be taken

into account to avoid movements. Still, some methods can be applied after

acquisition to minimize these movement artifacts. These methods assume that the


brain does not change in shape during the experiment and that any motion can be

corrected for by rotations and translation along x , y , z axes. This is referred to as

rigid body transformation with six degrees of freedom (3 rotations and 3 translations)

[19]. The rotation (Rr

) and translation (Tr

) matrixes are defined based on a reference

volume (usually the first one):

TxRxrrrr

+= )1()2( (1.11)

where )1(xr

and )2(xr

are the positions of the voxel vectors before and after the

correction and the matrixes Rr

and Tr

(Figure 1.4) are found to be the ones to

minimize the sum of squared difference between two voxels of subsequent volumes.

Abrupt motion can lead to artifacts in time series even after motion correction,

thus data with motion exceeding 2 mm are usually discarded.

Figure 1.4. Matrixes Rr(rotation) and T

r(translation) found during the motion correction

preprocessing step. Graph created using SPM software version 5 (www.fil.ion.ucl.ac.uk/spm).

Normalization

Until now, it was described how to identify activated brain areas in a single

subject. In order to identify the exact anatomical location, and to generalize such

individual findings to a population level and to make results from different studies

comparable, it is necessary to standardize data to a normative space. The


normalization step in the statistical parametric mapping (SPM) software registers

images of different subjects into the same atlas based coordinate system defined by

the Montreal neurological institute (MNI) using a large series of MRI scans on healthy

subjects. The registration consists of linear (accounts for major differences in head

shape and position) and nonlinear transformations (accounts for smaller-scale

anatomical differences). The linear transformation determines 12 affine

transformations (3 translations, 3 rotation, 3 shears and 3 zooms) to match positions

and sizes of images (Figure 1.5).

Figure 1.5. Visual representation of linear and nonlinear transformations held in the SPM

software normalization step. Figure taken from [20].

In order to increase the robustness and accuracy of the approach, information

about the variability of head sizes is included within a Bayesian structure. Still, given

the small number of parameters, it does not allow for every feature to be matched


exactly. Warps, therefore, are modelled by linear combination of smooth discrete

cosine transformation basis functions to correct gross differences in head shapes that

cannot be accounted for by the affine transformation [21].

A well know and frequently used atlas is the Brodmann atlas which was

created based on the cytoarchitectural organization of the human cortex by a

German neurologist, Korbinian Brodmann [22]. Later on, many of the Brodmann

areas could be associated with diverse cortical functions (Figure 1.6).

Figure 1.6. Brodmann atlas and common associated function. Figure taken from [23].

Spatial smoothing

The main goal of spatial smoothing is to suppress noise while maintaining the

signal of interest. Another reason to apply spatial smoothing is that the subsequent

statistical analysis requires the data to be normally distributed and spatial smoothing


of the data has a normalizing effect [19]. Individual differences in brain anatomy also

become less pronounced by smoothing and hence a cross-subject analysis is more

valid. The spatial smoothing is usually carried out by convolving a 3D volume with a

Gaussian kernel. The smoothing filter is defined by its full width half maximum

(FWHM), referring to the width of the Gaussian curve at half of its maximum.

Temporal filtering

The temporal filtering removes components of the temporal series that

fluctuates faster (lowpass filter), like pulsatil rhythms and frequency above the

Nyquist limit, or slower (highpass filter), like MR scanner drifts, than the predicted

activation. Different from the spatial filtering that is applied in each volume separately,

the temporal filtering is applied in the time series of each voxel. As statistical

analyses are directly applied on the time series, this filtering should be performed

after all the other preprocessing steps.

Statistical analysis

SPM has been extensively used to characterize neuroimaging data

sequences. It uses the general linear model (GLM) and classical statistics, under

parametric assumptions, to perform a statistical test (usually the T-statistics) at each

voxel. Inferences about regionally specific effects are based on the resulting image of

T-statistics (Figure 1.3b) [24]. The GLM describes the recorded data ( y ) as the

combination of weighted functions representing the expected HRFs (predictors) plus

an error (e ):

eXXXy LL ++++= ββββ L22110 , (1.12)


where iβ is the weight of the predictor function iX , 0β represents constant factors,

e.g. the baseline fMRI image intensity, and L the total number of predictor functions

needed to describe the experiment. For an experiment of J recorded volumes,

Equation (1.12) can be written in a matrix notation as:

eXY

e

e

xx

xx

y

y

JLJLJ

L

J

rrrr

MM

L

MOM

L

M

+×=

+

=

β

β

β 11

1

1111

(1.13)

Yr

is the column vector of observed data, er

and βr

are the error and weight vectors

respectively. Xr

is the LJ × matrix with each line representing one recorded volume

and each column a predictor function. The predictor functions are created with the

assumption that the BOLD signal changes can be estimated by convolving the HRF

with the stimuli onset times. Figure 1.7 shows a schematic representation of the

GLM.

Figure 1.7. Schematic representation of the general linear model (GLM). The paradigm

consisted of an auditory stimulation (red blocks) intercalated by a tapping finger task (blue

blocks) and a rest period (white blocks). The prediction functions were supposed to be the two


stimuli presentation time course convolving with the HRF ( 1x and 2x ). The two last time

courses are examples of possible measured time course of two different voxels in the fMRI.

The insert on the lower left shows typical activation maps derived by SPM showing activation

in the auditory cortex associated with tone presentation (top) or in the motor cortex

representing the finger taping condition (bottom).

βr

values are estimated by minimizing the residual values ( 0→er

), thus if

( )Lβββ~

,,~~

1 L= are the estimated values and ( ) β~~

,,~~1 XYYY J

rL == the adjusted values,

then ( ) β~~

,,1 XYYYeee j

rK

r−=−== . The residual sum of squared will be

eeeS TJ

j

j ==∑=1

2 , that is the sum of squared difference between the measured values

and the estimated ones. Finding the minimum of er

implies 0~ =∂

∂

l

S

β and this leads to

( ) YXXX TT 1ˆ −=β (if ( )XX T is invertible) where β is the optimized value. These

simple equations can be used to implement a vast range of statistical analyses [25].

Model free analysis

More recently, another method of fMRI analysis has emerged, focusing on

spontaneous fluctuations or “intrinsic activity” of the brain instead of evoked activity.

The brain at rest is responsible for most of the cerebral energy consumption whereas

task related increases in neuronal metabolism are usually less than 5%. The term

“intrinsic activity” summarizes any ongoing neural and metabolic activity that is not

directly associated with the subject performing a task. It is usually recorded when the

subject is asked to relax but not fall asleep in the scanner with eyes closed (resting-

state).


How can these intrinsic brain processes be analyzed [26]? One way to do this

is to apply an independent component analysis (ICA). ICA performs hypothesis-free

separation of input data into the linear sum of time varying modulation of maximally

independent component maps. It does not need any a priori assumptions about the

time course of the different component activations, e.g. whether a given component

is active due to specific neuronal systems or is related to technical noise or other

artifact sources [27].

It is acknowledged that the brain represents a complex network of dynamic

systems, consisting of numerous functional interactions between various brain

regions [28]. In this regard, fMRI can also be used to study such functional

connectivity, e.g. temporal dependency of neuronal activity patterns in anatomically

distinct brain regions. The evaluation of these dependencies can contribute to a

better understanding of the brain’s functional organization. Several approaches have

been proposed to quantify functional connectivity patterns [29]. In Chapter 2,

functional connectivity analysis is used on resting-state data to assess the

connectivity of the hippocampal formation with the neocortex during different sleep

stages. The time-course of a region-of-interest (ROI) is simply correlated against the

time course of all other voxels and a high correlation means high functional

connectivity.

Chapter 1 32

Electroencephalography (EEG)

The communication processes between the 100 billion neurons in the human

brain generates electrical activity, part of which can be recorded by the EEG. The first

EEG data were recorded in 1924 by a German psychiatrist, Hans Berger, and

reported in the essay "Electroencephalogram in Humans” in 1929 [30]. Since then,

this technology has continuously improved and it still widely used to study intrinsic

and evoked brain activities.

All neurons are covered by membranes that isolate and separate different

concentrations of ions inside and outside the cell. Through the neuronal membrane

are pore-forming proteins (ions channels) that manage the flux of specific ions in and

out of the cells. Here, the electrochemical driven force is determined by the electrical

potential difference across the membrane and the concentration gradient of the ions

selective for the channel. In the inactive state of the cell, this force creates a potential

difference across neural membrane, called membrane potential [31]. In this way, two

main forms of neuronal activation can be distinguished; the fast depolarization of the

neural membranes, which results in the action potential mediated by the sodium and

potassium voltage-dependent ionic conductance, and the slower changes in the

membrane potential due to synaptic activation, chemically mediated by several

neurotransmitter systems [32]. The electrical activity as recorded in the EEG is a

measure of the extracellular current flow from the summed activity of many neurons.

The scalp EEG mostly reflects the activity of cortical neurons close to brain surface.

Deeper structures as the hippocampus, thalamus or brainstem do not contribute

directly to the EEG. Distant transmission of electrical impulses, however, has

substantial effects on the surface EEG. Thalamocortical connections, for example,

Chapter 1 – Electroencephalography 33

are critical in the synchronization of most electrical activity, such as sleep spindles

[31]. EEG scalp events can therefore (indirectly) reflect more deep-brain

activity/connectivity.

Only vertically oriented dipoles are detectable with scalp electrodes. The part

of the cortex that runs down the sulci, oriented orthogonal to the scalp, generate

radially oriented dipoles, which are not well detected by the surface electrodes. Thus,

the activities recorded by scalp electrodes are mostly generated by excitatory

postsynaptic potentials (EPSP) produced in the pyramidal layer of the cerebral

cortex. The pyramidal neurons are spatially aligned and perpendicular to the cortical

surface and when activated, with a certain degree of synchrony, generate coherent

electric/ magnetic fields that can be detected by electrodes placed at relatively small

distances. [32]. Figure 1.8 illustrates transient dipoles that lead to measurable

extracellular voltage.

Figure 1.8 Afferent inputs into the apical (a) dendrites and (b) cell body. In both cases, the

afferent stimuli lead to depolarization with current flow into the cell body. The current flow in

(a) results in a source in the apical dendrite, whereas in (b), the source is located in the soma.

This example thereby leads to two vertically oriented dipoles of opposing polarity that can be


detected by surface EEG. Figure reproduced with kind permission from Springer

Science+Business Media: Basic Neurophysiology and the Cortical Basis of EEG, 2007, page

31, Gregory L. Holmes and Roustem Khazipov, figure number 6, Modified from Kandel et al.,

2000 with permission [31].

Unfortunately, EEG carries the problem that signals measured on the scalp

surface do not directly indicate the location of the active neurons in the brain. Many

different source configurations can generate the same distribution of potentials on the

scalp. This ambiguity of the underlying static electromagnetic inverse problem has no

unique mathematical solution [33]. The only way to localize the putative electric

sources is to introduce a priori assumptions on the generation of the EEG signal. In

the last years, many assumptions using different mathematical, biophysical,

statistical, anatomical and functional constraints have been formulated and

implemented in inverse solution algorithms [33]. Basically, there are two main

approaches for estimating sources from observed sensor EEG: The first assumes

that sensor data can be explained by a small set of equivalent current dipoles.

Secondly, the source reconstruction problem has more recently been modeled by

placing many dipoles in brain space and using certain constraints on the solution to

make it unique. Most of these constraints are anatomical and physiological

arguments, e.g., smoothness constraints and approximate location priors based on

regional activity in fMRI.

EEG acquisition and preprocessing

In a typical EEG experiment, electrodes are positioned at standardized

locations on the scalp, as for example defined in the international 10-20 system. The

"10" and "20" refer to the distances between adjacent electrodes that are either 10%


or 20% of the total front-back or right-left distance based on reference points on the

skull, respectively. This method was developed to ensure experimental reproducibility

and to allow findings to be compared over time (see Figure 1.9).

Figure 1.9. The international 10-20 system seen from (a) left and (b) above the head. A = Ear

lobe, C = central, Pg = nasopharyngeal, P = parietal, F = frontal, Fp = frontal polar, O = occipital.

Figure was taken from [34]. (c) Location and nomenclature of the electrodes used in the

present study. AFz and FCz are the ground and reference electrode. (d) Visualization of typical

caps used in EEG studies. Figure 1.9a and 1.9b were taken from [34]. Figures 1.9c and 1.9d

were taken from [35].

Beside the mathematical and statistical approach used to interpret the EEG

data, technically satisfactory recordings are required to ensure accurate results. EEG

is more difficult to measure due to its low amplitude (in the µV range) caused by the

skull’s composition if compared with other noninvasive biosignal measurements such

as the electrocardiogram (ECG), electromyogram (EMG) and electrooculogram

(EOG). The first challenge is to thoroughly place the electrodes on the subject’s

head. The skin and deeper tissues between a pair of electrodes have resistive and

capacitative properties. The impedance of these tissues can be decreased by


preparation of the skin using suitable abrasion underneath the electrode contacts. In

a noninvasive electrical brain signal measurement, an interface material is applied

between the electrode and the skin to provide optimal ionic current and to reduce

contact impedance between the electrode surface and the scalp. This material is an

electrolyte and can be purchased in EEG gel or paste form.

During recording, technical artifacts can often appear due to loose contact of

electrodes, which results in electrical noise or baseline drifts. The baseline drift

artifact can be seen as another type of electrode artifact. The power spectrum of the

low frequency component increases sharply when the baseline drift occurs. An

unstable impedance of the electrode is the main reason leading to such artifacts, but

sweating and body movement may also add to this effect [36]. To compensate for

baseline drifts, highpass filters can be applied. Other off-line correction of DC drifts

artifacts was proposed by Hennighausen in 1993 [37]. The main idea of this method

is to calculate the average of consecutive EEG segments and perform a regression

of these values. The amplitude trend which can be explained by either a linear or

non-linear regression model is then subtracted from all data points. It should not be

forgotten, however, that slow drifts could also represent meaningful activity related to

habituation, attention or arousal level of the subject during the experiment.

EEG recordings are contaminated by physiological artifacts as eye blinks or

other eye movements, as well as by EMG and ECG activity. The movement coming

from eyes or muscles cause rather large signals, which may obscure the low

amplitude EEG signals coming from the brain [36]. Many of these artifacts cannot be

corrected for, and the EEG segments in which they occur must be excluded from

analysis. One frequently used attempt to correct for ECG and EOG artifacts is


applying ICA. Components with strong ECG and EOG related time courses are

visually identified and removed from the data after an ICA back-transform.

Chapter 1 38

Simultaneous EEG and fMRI recordings

EEG/fMRI simultaneous data acquisition is becoming more common and

essential in both basic and clinical brain research. Cognitive neuroscience,

epileptology, psychiatry and sleep research all benefit from the combination of these

methods. This multimodal approach is a key tool to enhance the understanding of

brain activity due to the conjunction of the good spatial resolution of fMRI and the

good temporal resolution of EEG, respect and its complementary aspect covering

electrophysiological and hemodynamic related changes. However, the successful

combination of EEG and fMRI demands careful consideration of patient safety, as

well as EEG and MR image quality.

The EEG equipment appropriated for use in the MR environment must

exclude ferromagnetic material and should limit rf emission to preserve image quality.

Potential sources of rf emission are active circuitry in EEG instrumentation located in

the scanner room or the ingress of rf signals via conductors that breach the scanner’s

Faraday shield. Rf signals should be minimized at source by using low-power digital

components, thereby minimizing switching currents. All active circuitry should be

enclosed in conductive enclosure and all conductive signal paths breaching this

enclosure should do so via inline rf filters. Normally, the EEG preamplifiers are

located in the scanner room and EEG data is transmitted to a receiver in the console

room via fiber optic cables, therefore eliminating the ingress of rf from outside the

scanner. Regarding the instrumentation material, fortunately, this is not a significant

limitation since there are a range of nonferromagnetic materials that are suitable to

high-quality EEG as silver, silver chloride, gold, carbon and conductive plastic EEG

electrodes [38].

Chapter 1 – Simultaneous EEG and fMRI recordings 39

The rapidly switching rf and magnetic gradient fields applied during MR data

acquisition induce eddy currents in the electrodes and may result in heating of the

electrode [39;40]. Several studies, however, showed that the local temperature

increase is less than 1oC in scanners of 1.5T and 4T [39;41;42], and thus within the

permitted limit [39]. The interaction between the scanner (rf) and the electrode leads

can result in tissue heating as well. The rf field will induce an electromotive force

(emf) in any conductive loop formed by the electrode leads which is proportional to

the loop area cut by the field and its rate of change. Thus, if tissue is part of this loop

the emf will drive a current through it, resulting in heating. The electric component of

the rf can induce a current along the extended conductor formed by an electrode

lead. The magnitude of this current depends on the wire to the source of electric field

in the MR transmitter coil, the resonant length of the electrode leads in relation to the

rf wavelength among others [38]. Lemieux et al. [39] compared both sources of

heating and found that the first one is stronger, but that a current-limiting resistor of

12 Ω in the scalp electrode leads could limit contact currents to acceptable levels.

This additional resistance is small compared with typical input impedance of an EEG

amplifier (around 10 MΩ) and does not degrade EEG signal quality significantly [38].

In the so performed EEG recording, the fMRI recording, including switching of

magnetic field gradients and transmission of rf pulses, induces several different types

of artifacts in the EEG, basically the gradient and the cardiac pulse-related artifacts.

The amplitude of the gradient artifact is in the range of 10000 µV (~100 times higher

than normal EEG) and can reach speeds of 20000 µV/ms (500 times faster than

spontaneous EEG), see Figure 1.10. The most frequently applied method to correct

this technical artifact comes from the assumption that two acquisition of the same


slice or volume will generate precisely the same artifact in the EEG data, thus the

artifact can be easily modeled in a template and subtracted from the raw EEG [43].

Figure 1.10 Example of gradient artifact in a concurrent EEG recording. (a) shows one ECG

(EKG1) and 13 EEG (F3, F4, C3, C4, P3, P4, O1, O2, T7, T8, Fz, Cz, Pz) channels. The ECG was

downscaled by a factor 20. (b) shows a zoom of the red box in (a) with strong gradient induced

artifact shadowing the underlying EEG. Time and amplitude are displayed for each. The marker

R128 is automatically delivered by the MR scanner and defines the exact time at which a new

slice acquisition starts. Note the high similarity of the gradient induced artifacts in (b).

The amplitude range covered and sampling rate of the EEG recorded inside a

MR scanner should be high (e.g. ± 16V, 5000Hz for BrainAmpMR) in order to detect

and correct the very fast artifacts introduced by the scanning procedures. Creation of

precise artifact templates in the EEG is much improved by exact knowledge on slice

acquisition starts. During recording, the scanner used in Chapter 2 and 3 (1.5 Tesla


Signa Echospeed, General Electric, Milwaukee, Wisconsin, USA) gave a TTL signal

output at the exact time it started a new slice acquisition. These signals were entered

and saved as markers in the EEG recording and can later be used to build a

correction template across all consecutive volumes (marker R128 in Figure 1.10).

Another prerequisite to a good template construction is that the onset of each fMRI

volume to be acquired coincides with the exact time point at which an EEG data point

is being sampled. Otherwise, there would be a temporal misalignment between the

two acquisition systems that invariably will lead to a massively increased variance or

error in the gradient correction template [43]. In order to achieve synchronicity

between an fMRI sequence and the regular sampling pattern of the concurrent EEG,

the repetition time TR of the fMRI sequence must be a multiple of the EEG sampling

interval and the internal clocks of both systems, EEG and MRI, must be phase

locked. Luckily, nowadays, commercial EEG recorders are equipped with

synchronization hardware that embraces this last issue [44]. Finally, after an optimal

data acquisition, a template based on an average of several imaging artifact

waveforms is calculated over a number of epochs and eventually subtracted from the

EEG [45]. As the high sampling rate is no longer necessary after gradient artifact

correction, and in order to decrease data size, the corrected data is down-sampled at

this point. In the software Brain Vision Analyzer, this down-sampling is performed by

using a Hanning window to calculate a weighted average for the data points. After

subtraction of the averaged artifact curve, low-pass and band-rejection filters may be

used to remove further artifacts.


Figure 1.11 Example of BCG artifact in a concurrent EEG recording. The figure shows one ECG

(EKG1) and 4 (C3, F7, T7, T8) EEG channel. Note that the amplitude resolution of the EEG

electrodes is different from the ECG. The red line helps to see that the BCG is slightly delayed

with respect to the ECG.

Another challenge is posed by the ballistocardiogram (BCG) artifact,

produced by cardiac pulse-related movements of the scalp electrodes inside the

static magnetic field. These movements might result from acceleration and abrupt

directional change in blood flow in the aortic arch during each heart beat [46]. Scalp

movement may also occur due the expansion and contraction of the scalp arteries

[47]. In an EEG channel, this artifact peak normally occurs slightly after the QRS

complex (delayed by ~ 200 ms) [47], see Figure 1.11. While gradient artifacts are

highly predictable, BCG artifacts display much greater variability due to their


physiological origin. The amplitude of the BCG can exceed 200 µV obscuring

completely the neural activity detected by EEG. Furthermore, it covers a substantial

frequency range coinciding directly with neural signals that are of interest, making the

use of filters to remove this artifact many times impractically [48]. Allen et al. [47]

introduced a method based on averaged artifact subtraction, which is characterized

by a subtraction of a mean BCG artifact template calculated for each electrode during

the previous 10 s. More recently, methods based on ICA were successfully used to

solve this problem [49;50]. As already mentioned, the strength of ICA is that it makes

no assumption about the mixing process of the different noise sources, except that

this occurs linearly. This method consists in recovering unobserved statistically

independent signals or “sources” from several observed mixtures or different

combination of the “source signals” [51].

Chapter 2 44

CHAPTER 2

Sleep spindles and hippocampal functional connectivity

in human NREM sleep

Introduction

Sleep

Sleep disturbances have a great impact on quality of life, morbidity/mortality,

public health and productivity [52] and can manifest in many psychiatric disorders like

depression and schizophrenia, among others [53-55]. During the last years,

neuroimaging has proven a valuable non-invasive tool to detect anatomical,

functional and metabolic changes associated with sleep disturbances. The beginning

of sleep research, however, is much older and is closely connected with the invention

of the EEG. Using EEG, in 1937, Loomis showed for the first time that sleep was not

homogeneous during the whole night but constitutes of different sleep stages [56].

Today, these sleep stages are scored according to the Rechtschaffen and Kales [57]

criteria based on EEG, EOG and EMG recordings. The main states of vigilance are

wakefulness, rapid eye movement (REM) sleep and NREM sleep. NREM sleep is

further subdivided into stages 1, 2, 3 and 4 with slow wave sleep (SWS) comprising

stages 3 and 4, see Figure 2.1. The rhythmic neuronal activity changes in sleep are

accompanied by changes in information processing and state of consciousness. The

transition from being fully awake to being clearly asleep is of great importance to

understand clinically relevant changes in alertness and sustained attention and is

associated with well-established EEG changes [58]. The EEG pattern associated with

each behavioral state is quantified by the power of specific rhythms in certain

frequency bands. Sleep stage 1 (S1) is characterized by a decrease in high

Chapter 2 - Introduction 45

frequency oscillations and a dominance of (slower) theta rhythms (4-8 Hz) in the EEG

and by lower EMG levels as compared with wakefulness and the appearance of slow

eye movements [59;60]. Behaviorally, it includes: a decrease sensory perception and

a partial cessation of responses to external stimuli [7;61;62]. S1 seems to be

associated with decreased activity in the frontal and parietal cortices, and in the

thalamus [10]. The thalamus plays a critical role in processing, integrating,

correlating, and relaying sensory and motor information The thalamocortical network

modulates the flow of sensory and motor information to and from the cerebral cortex

[62]. Starting in early NREM sleep S1, the thalamocortical neurons undergo

progressive hyperpolarization leading to a reduced synaptic responsiveness and

interruption of the bidirectional flow of information. This has been further linked to

sleep spindles (Figure 2.1b) [62-66] during deeper NREM sleep stages. Sleep

spindles are waxing – and – waning 11 - 15 Hz oscillations and can be subdivided in

slow (11 - 13 Hz) and fast (13 - 15 Hz) spindles, predominant in frontal and centro-

parietal areas, respectively [64]. Sleep spindles are most prevalent during sleep

stage S2, which is also associated with a low level of consciousness. S2 is also

characterized by the presence of a sharp negative wave followed by a slower positive

wave pattern, called K-complex (KC), see Figure 2.1b. KCs can appear

spontaneously or be evoked by external stimuli. SWS is associated with the deepest

level of unconsciousness and lowest energy metabolism and is dominated by a high

amplitude EEG with frequencies lower than 4 Hz [59;67;68]. REM sleep, also called

paradoxical sleep, shows an EEG resembling wakefulness or S1 sleep, and is further

characterized by atonia of voluntary muscles (e.g. not affecting heart or respiratory

muscles, and in addition eye muscles) as reflected by lowest chin EMG level as well


as by rapid eye movements (Figure 2.1c) [57;69]. Awaking from this state of high

cortical activity goes along with higher incidences of vivid dream reports [63;67].

Figure 2.1. Relationship between human sleep, level of consciousness and EEG patterns. (a)

Stages of sleep are characterized by differences in the frequency and amplitude of EEG waves.

Wakefulness is characterized by a desynchronized high-frequency EEG. Stage 1 comprises

light sleep with low-amplitude waveforms and slightly decreased frequency. Stage 2 is

characterized by sleep spindles (higher-frequency waves around 11 Hz – 15 Hz) and K-

complexes (graphically represented in (b)). Stages 3 and 4 comprise SWS with high-amplitude

low-frequency waves. These four stages constitute NREM sleep. Another sleep stage is the

REM sleep, which is associated with theta activity (4.5-7.5 Hz) and, as its name implies, is

characterized by rapid eye movements (graphically represented in (c) by two horizontal EOG

channels of the left and right eye). REM is also associated with low-amplitude sawtooth waves

(not shown) and a higher level of consciousness. Figure 2.1 (a) was reproduced by permission

from Macmillan Publishers Ltd: Nature Reviews [67], copyright (2004). Figure 2.1 (b) was

reproduced from [59]

There are many theories which try to explain functions and the purpose of

sleep. The main theories are: conservation of energy, restoration of tissue and


growth, thermoregulation, regulation of emotions, neuronal maturation, and memory

and learning [59;70;71]. A full discussion of these theories goes beyond the scope of

this thesis.

Memory and Hippocampus

Among the many functions attributed to sleep, its role in the memory

consolidation is of particular importance as sleep may be required to redistribute and

reactivate recent memory traces in the absence of interfering external inputs [71-74].

Whereas memory encoding and memory retrieval is bound to wakefulness, sleep has

been postulated to promote memory consolidation. Consolidation refers to the

process that transforms new and labile memories collected during wakefulness into a

more stable cortical representation, making them integrated in networks of pre-

existing long-term memories [75]. Human memory can be subdivided in declarative

and nondeclarative memory. Declarative memory summarizes memories that are

consciously accessible, such as fact and events (knowing “what”, e.g. “what is the

capital of Brazil?”). Current neuronal models show the importance of the temporal

lobe, including the hippocampus, in declarative memory formation. Nondeclarative

memory includes the knowing “how” (procedural memory), for example actions,

habits and skills (e.g. how to ride a bike); it is less dependent on medial temporal

lobe structures [76]. Several studies support the idea that declarative, hippocampus-

dependent memory is particularly strengthened by SWS, whereas nondeclarative,

procedural memories benefit to a greater extent from REM sleep [77-79]. However,

controversial results were also reported in which SWS can also improve

nondeclarative, as well as REM sleep can improve declarative memory [80-82]. S2

has been associated with both kinds of memory consolidation [83-85]. Schabus et al.

[86] measured spindle activity during S2 following a declarative memory task and a


control task. They found that increased S2 spindle activity was related to an increase

in recall performance, reflecting memory consolidation. It is important to note that

most memory tasks are neither purely declarative nor only procedural, and that both

systems interact during sleep-dependent memory consolidation.

According to the standard model of memory consolidation at the systems-

level, active cerebral communication transfers new memories, initially encoded in the

hippocampus, into long-term memory representations stored in the neocortex [87-91].

Thus, remote memories are based on neocortical networks and can be retrieved

independently of the hippocampus [92]. The (re-)activation of the hippocampus is

reflected by hippocampal ripple activity that during sleep is found synchronized to

sleep spindles [93; 94]. Thus, the hippocampus has been proposed to orchestrate

the reactivation of memory traces and their reinstatement in cortical circuits [95]. The

hippocampal-cortical system includes the hippocampal formation as well as its

widespread cortical targets. The hippocampal formation (HF) consists of the Cornu

ammonis (CA), dentate gyrus (DG), and subiculum (SUB) that are forming as a loop-

like structure and are receiving and projecting information to the neocortex. This

information flow is passing through the adjacent entorhinal cortex (EC) [96], see

Figure 2.2. Next to the function of the HF related to memory processes, it has also

been associated with spatial encoding, inhibition and anxiety [98;99]. Less clear is

the precise connectivity of the hippocampus with the neocortex, not only anatomically

but also in terms of effective functional connections. In this context, most knowledge

stems from animal models [100;101].


Figure 2.2. Cytoarchitecture of the mesial temporal lobe at the level of the body of the

hippocampus. The hippocampal subregions and entorhinal cortex are displayed with different

colors. Reproduced with kind permission from Springer Science+Business Media:

Cytoarchitectonic mapping of the human amygdala, hippocampal region and entorhinal cortex:

intersubject variability and probability maps, 2010, 2005, pag 345, K. Amunts et al, Fig. 2 [97].

Resting-state

In fMRI, spontaneous signal fluctuations have recently been used to study

functional cerebral connectivity [102]. Such signal fluctuations are organized in

distinct functional resting-state networks (RSNs). As mentioned in Chapter 1, they

are usually recorded in subjects asked to relax but not fall asleep in the scanner, with

eyes closed. Figure 2.3 exemplifies some of these RSNs.

Among the various RSNs, the “default mode” network (DMN) received

particular attention, being linked to self-referential processes in humans, such as

autobiographical memory and future envisioning [104]. During wakefulness it consists

of subsystems in which the hippocampal formation and the lateral temporal cortex

are involved in memory collection and building of associations, flexible use of this

information to construct mental simulations, and integration of these processes


[96;104]. Also the posterior nodes of the DMN (posterior cingulate, precuneus, and

inferior parietal lobules) correspond to regions involved in memory recollection [105].

Figure 2.3. Different resting-state networks: RSNs were identified using ICA that decomposed

the individual’s fMRI into 30 statistically independent components (ICs), each delineating

groups of voxels that exhibited synchronous temporal fluctuations. (a) “Default mode”, (b)

frontoparietal control, (c) frontal attention, (d) somato-motor, (e) auditory and (f) occipital

visual network. Color bar depicts T-values. RSNs taken with permission from Jann K, Kottlow

M, Dierks T, Boesch C, Koenig T (2010) Topographic Electrophysiological Signatures of fMRI

Resting State Networks. PLoS ONE 5(9): e12945, Fig 1 [103].

Preserved DMN connectivity was reported in light sleep [106;107], while the

DMN showed disintegration in slow wave sleep [107;108]. More subtle network

changes, like a reduced contribution of the HF to the DMN, have also been shown at

sleep onset [108]. However, as these studies focused on the DMN as a whole entity,

it remains unclear if the HF builds up alternative functional connectivities with brain

areas outside the DMN throughout the course of sleep.


Objective and hypotheses

The main goal of this work was to investigate the functional connectivity of the

hippocampal subregions (CA, DG and SUB) to all other brain regions for wakefulness

(W), and the NREM sleep stages 1 (S1), 2 (S2) and slow wave sleep (SWS), by

means of simultaneous EEG/fMRI data collected during sleep.

The a priori hypotheses were:

1. The HF shows strongest functional connectivity to the major DMN

nodes in wakefulness.

2. In stable sleep stages (S2 and SWS), different functional connectivity

should be revealed. Based on the assumption of increased hippocampus-to-

neocortex transfer, frontal and temporal brain regions should be more involved.

3. Functional connectivity changes during sleep are postulated to be

associated with spindle activity as derived from concurrent EEG recordings, based on

the assumed relation of hippocampal ripple and sleep spindle.

4. Regarding the role of hippocampal subregions (dentate gyrus and

cornu ammonis) and the subiculum, we hypothesize especially enhanced functional

connectivity of the hippocampal output region, the subiculum, during sleep.

Chapter 2 52

Methods

Subjects

The study protocol was in accordance with the Declaration of Helsinki and

was approved by the ethical review board of Ludwig-Maximilians University, Munich,

Germany. Participants provided their written informed consent after the procedure

had been fully explained and were reimbursed for their participation. Subjects

underwent a general medical and structured psychiatric interview, clinical MRI and

EEG to exclude clinical conditions that could interfere with the study protocol.

Combined fMRI/polysomnography was obtained from 25 healthy adults (13 men,

mean [SD] age 24.7 [3.2] years, 12 women, 24.8 [2.5] years). Participants were

instructed to follow a regular sleep-wake-schedule with bedtimes between 23:00 and

08:00 hrs during the week prior to the experiment, documented by sleep diaries.

Participants were asked to get up about three hours earlier on the experimental day

to increase the probability of falling asleep in the MRI scanner. Wrist actigraphy

during the night and day before the experiment was employed to control subjects’

adherence to the sleep restriction. Simultaneous EEG/fMRI experiments started

around 9:00 pm. After EEG montage and positioning in the MRI scanner subjects

were informed that no further active participation was required during the following 2-

3 hours, and that they could fall asleep.

FMRI and EEG acquisition

Simultaneous polysomnography comprised 19 EEG electrodes placed

according to the international standard 10/20 system, electrooculogram, submental

electromyography and an electrocardiogram (sampling rate 5 kHz; EasyCAP

Chapter 2 - Methods 53

modified for sleep, Herrsching, Germany; VisionRecorder Version 1.03,

BrainProducts, Gilching, Germany). For noise protection subjects wore ear plugs and

headphones. FMRI was carried out at 1.5 Tesla (Signa LX, GE, Milwaukee, USA)

using an 8-channel head coil.

One fMRI run consisted of 800 functional whole brain images (EPI; repetition

time 2000 ms; echo time 40 ms; flip angle 90°; 64 × 64 matrix, in-plane resolution 3.4

× 3.4 mm2; slices 25; slice thickness 3 mm; gap 1 mm; oriented along AC-PC)

acquired over 26.7 minutes. This time restriction originated in a scanner software

limitation to continuous acquisition of 20000 slices at maximum. To allow for vigilance

monitoring during the recording session, the sleep stages were determined online

using the real-time gradient EEG artifact correction of the Vision Recorder (Version

1.03.0003) and Vision RecView (Version 1.0) software (Brain Products, Gilching,

Germany). If the subject was not able to fall asleep, only reached light NREM sleep

stages or showed an early and rapid transition to S2 and SWS during the first run,

acquisition was repeated to increase the probability of recording the missing sleep

stages. In total, 40 fMRI runs of each 26.7 minutes length were acquired from 25

subjects.

Sleep stage rating

EEG preprocessing was performed using Brain Vision Analyzer software 1.05

(Brain Products, Gilching, Germany). Correction of MRI related EEG artifact was

performed by subtraction of an adaptive template based on artifact averages of 8

consecutives MRI volumes. Independent component analysis was applied for

removal of ECG and EOG artifacts. Components with strong ECG and EOG related

time courses were visually identified and removed (see Chapter 1 for more details


about EEG artifact correction). Data were resampled to 250 Hz, re-referenced

against linked mastoids and filtered (0.1 Hz to 30 Hz). Sleep stage scoring following

Rechtschaffen and Kales [57] criteria of all 40 fMRI runs was performed in 20 second

windows. The hypnograms of the 40 recorded 26.7-minute sessions were then

screened for intervals consisting of 5 minutes of one specific, stable sleep stage (due

to some waxing and waning of sleep defined by presence of more than 85%

throughout the epoch), without arousals or movement artifacts in either EEG or fMRI.

This resulted in 93 epochs of the following vigilance stages: wakefulness (W: 27),

sleep stage 1 (S1: 24), sleep stage 2 (S2: 24), and slow wave sleep (SW: 18).

FMRI analysis

1. Preprocessing: All fMRI analyses were performed in 64-bit Linux

workstations using SPM, version 5 (www.fil.ion.ucl.ac.uk/spm), and in-house scripts

programmed in IDL version 6.3 (www.ittvis.com) and Matlab version 2008B (The

MathWorks, Natick, USA). The fMRI preprocessing steps consisted of: 1) slice time

correction to account for different acquisition times between slices; 2) realignment

using rigid body transformation to correct for head motion. Data with head

movements more than 2 mm were excluded; 3) spatial normalization of the images to

a standard EPI template in MNI space (SPM5 distribution) and interpolation to a

voxel resolution of 2 × 2 × 2 mm3 using 5th degree splines (see Chapter 1 for fMRI

preprocessing details).

2. Removal of artificial signals: To remove artificial components and signal

contributions of non-neural origin from the time courses, spatially normalized and

unsmoothed images were residualized by multiple regression analysis in the GLM

framework against the following regressors: (i-vi) six parameters derived from the


motion correction step (matrixes Rr

and Tr

), (vii-xii) their respective first order

derivatives, (xiii-xiv) the averaged time series of the white matter and CSF voxels,

(xv-xvi) and their respective first order derivatives. They resultant in residual images

contained signal fluctuations not explained by regressors i–xvi. They were temporally

(lowpass filter of 0.1 Hz in FSL 3.2 (www.fmrib.ox.ac.uk/fsl)) and spatially (isotropic

Gaussian kernel, full width half maximum of 6 mm) filtered and afterwards used for

extraction of region-based time courses.

The 0.1 Hz lowpass filtering was used to restrict analysis to the typical

frequency range generally used to study resting-state network fluctuations due to the

expected functional connectivity of the hippocampus and DMN. Such restriction may

shadow possible higher frequency contributions to spontaneous signal fluctuation. To

test for the effect of lowpass filtering, the analysis has also been performed without

filtering (results not shown), leading to the same quantitative results. However,

physiological noise contributes much more to the later analysis, and therefore the

lowpass filtered data was in the focus of this work.

3. Fixed effects model: For each subject and sleep stage, data were

regressed to the corresponding hippocampal time courses. Subregions of the

hippocampal formation were defined from probabilistic cytoarchitectonic maps as

previously described [97] (Figure 2.4) and implemented in the SPM anatomy toolbox

[109]. Time courses were calculated as the mean signal of all voxels inside the

specific region, and extracted using the Marsbar toolbox [110] for SPM. The following

hippocampal subregions were used: Cornu ammonis (CA), dentate gyrus (DG), and

subiculum (SUB). For every subject and sleep stage, each time course of the bilateral

subregion was regressed on the residual fMRI images in three separate fixed effects

analyses.


Figure 2.4. Probabilistic cytoarchitectonic maps of the hippocampal subregions and subiculum

as previously described in [97]. Red - cornu ammonis, yellow - dentate gyrus and green –

subiculum.

4. Random effects model: For analysing sleep stage influences, a second

level random effects analysis was performed using a two-factorial design (factor

subregion: three subregions, factor sleep: wakefulness and three NREM sleep

stages). A map of the HF functional connectivity during wakefulness was generated

by a t-test against zero combining all subregions. Effects of factor sleep and of factor

subregion were tested using analysis of variance (ANOVA). Directed t-tests were

performed in addition to explore the direction of the effects (W vs S1, W vs S2, W vs

SWS, S1 vs S2, S1 vs SWS, and S2 vs SWS). Finally, specific subregional

contributions were identified per sleep stage by contrasting each subregion against

the two other subregions combined.

The functional connectivity map collected during wakefulness was

thresholded at pFWE<10-6, extent > 150 voxels. For the main effects of factors sleep

and subregion, for the bidirectional sleep stage comparison as well as for the

subregional analyses, an uncorrected threshold was set at p<0.001, with varying

cluster extent in order to ensure significance of the resulting clusters under

consideration of non-stationary smoothness (pFWE,cluster<0.05) [111].


Sleep spindle analysis

Analysis of sleep spindles and related fMRI activation was based on the

procedures described in Schabus et al. [64]. In the EEG data of the selected epochs

in sleep stage 2, occurrence of slow and fast sleep spindles was determined after

band-pass filtering of 11 - 13 Hz (0.0145 s, 48 dB/oct) and 13 - 15 Hz (0.0122 s, 48

dB/oct), respectively. The root mean square (rms) of the filtered signal was calculated

in time windows of 200 ms. For slow spindles, EEG data collected at electrode Fz

were used, as slow spindles are known to be strongest over frontal brain regions,

while for fast spindles (with parietal dominance) Pz was used [112]. For both spindle

types, the 20% highest rms amplitudes were classified as spindles. Finally, onset of

the classified spindles were used to construct the event regressor and later

convolved with the three basis canonical function (hemodynamic response function,

its time and dispersion derivatives). As in Schabus et al. [64] the power fluctuation in

the EEG delta band (0.5 – 4 Hz) was added as a covariate, as spindles are

modulated by slow-waves. For this purpose, we extracted the delta power from Fz

per fMRI volume using Fast-Fourier Transform (FFT, resolution = 0.5 Hz, Hanning

window 10%), and convolved it with the hemodynamic response function. For each

subject, the event regressors of fast and slow spindles convolved with the three basis

canonical functions along with delta power were regressed against the residual fMRI

images in a fixed effects model.

Second level analysis was performed using the individual t-contrasts of each

of the three canonical basis functions as well as of the two spindle types as factors.

The error covariance was not assumed independent between regressors and a

correction for non-sphericity was applied [113]. The resulting maps were collected at

puncorr <0.001, extent > 135 voxels, signifying pFWE,cluster < 0.05 [111].


Psychophysiological interaction analysis

Psychophysiological interactions (PPI), refers to a method relating the

functional coupling between one area and the rest of the brain to results of

psychological tests. Here, as external variable an electrophysiological, not

psychophysiological means was applied, however recorded simultaneously and using

the same methodological approach. PPIs between the EEG spindle data and the

BOLD signal time courses of the three HF subregions were evaluated in order to test

if changes in HF functional connectivity had an interaction with increased spindle

activity. PPIs were performed for CA, DG and SUB in three distinct analyses. For

each subregion, first, the neuronal representation of the signal time course of the

subregion was approximated by deconvoluting the HRF from it. Second, the neural

representation of the fMRI signal time course was convolved with the individual event

regressor containing the onset time points of spindles (also one analysis for each

spindle type, slow and fast spindle) and afterwards convolved with the HRF. The

regressor which resulted from this transformation and the two original regressors

before the convolution (fMRI signal time course of the subregion and HRF folded

event regressor containing the onset time points of spindles) were entered in a fixed

effect analysis. Analysis was performed on the residualized data as described above.

Data were thresholded at puncorr<10-4, extent k>150, pFWE,cluster < 0.005.

Chapter 2 59

Results

Subjects adhered to the experimental guidelines and sleep restriction, as

shown by sleep diaries and wrist actigraphy. Sleep latency (time until first

appearance of S2), only considering the first attempt of the subject to fall asleep, was

6.7 ± 4.9 minutes. Over all 40 fMRI runs, the average time spent in W, S1, S2 and

SWS were 6.8 ± 5.2 minutes, 5.9 ± 4.2 minutes, 10.0 ± 5.4 minutes and 4.8 ± 5.8

minutes, respectively. Eventually, a closer inspection of hypnograms identified 93

epochs (27, 24, 24 and 18 epochs of W, S1, S2 and SWS of 15, 18, 10 and 11

subjects) of each five contiguous minutes with a single dominating vigilance stage,

i.e. less than 14 % of time spent in flanking sleep stages and without arousals.

Assignment of all 5-minute epochs to subjects and runs including a description of the

prevalence of the different sleep stages per run is detailed elsewhere [114].

Occurrence of REM sleep was not observed, which is not unexpected as data were

obtained only from the initial part of the sleep cycle and REM sleep is suppressed in

the noisy MR environment [63]. During later processing, one data set acquired during

S1 was found to be corrupted and excluded.

Sleep stage and HF functional connectivity

Figure 2.5a shows the functional connectivity map for HF during

wakefulness. HF is functionally connected, amongst others, to the major nodes of the

DMN, namely the posterior cingulate cortex/retrosplenial cortex (PCC/RspC), medial

prefrontal cortex (mPFC), bilateral inferior parietal lobule (IPL) and bilateral middle

and superior temporal gyrus, as well as the thalamus (Table A2.1, Appendix A2).

The HF formation is most strongly connected to the DMN during wakefulness, with

Chapter 2 - Results 60

gradually decreasing functional connectivity with the DMN in NREM sleep stages, as

illustrated in Figure 2.5b for the DMN core regions.

Figure 2.5. (a) Extension of the HF network during wakefulness. Positive functional

connectivity of DMN nodes and the HF is shown (pFWE<10-6, extent>150 voxel). Color coding

indicates T-values. MNI coordinates of each slice are given. (b) Contrast estimates ± SD,

deviations extracted at the peak voxel of the indicated cluster and separated for each sleep

stage (wakefulness: W; sleep stages 1 and 2: S1/S2; slow wave sleep: SWS), are shown. MNI

coordinates (x,y,z) of the cluster peak voxel are provided. PCC: Posterior cingulate gyrus;

mPFC: Medial prefrontal gyrus; IPL: Inferior parietal lobule.

To further examine sleep stage specific alterations in HF functional

connectivity, an F-test was performed on the factor sleep (Figure 2.6).

Figure 2.6. Results of the full-factorial design employing factors sleep stage and subregion. F-

test showing the main effect of sleep. Data are thresholded at pFWE,cluster<0.05 using a collection

threshold of p<0.001. Note that the HF has not been masked out. MNI coordinates of each slice

are given. Color bar depicts t-values.


Figure 2.7 Shows the directed t-tests performed to explore the direction of

the changes comparing wakefulness against all NREM sleep stages, as well as for

the comparison between S2 and SWS (Table A2.2, Appendix A2). Comparing

wakefulness with S1, only the thalamus showed higher functional connectivity with

the HF during wakefulness. During S1, functional connectivity of the HF was

generally found to be stronger in the middle and superior temporal gyrus, occipital

cortex and inferior parietal lobule (Figure 2.7a). During S2, HF functional connectivity

to the lateral temporal and the occipital cortex further increased as compared with

wakefulness, while no differences in thalamic connectivity were visible. In addition,

during S2 functional connectivity of the HF with the cingulate cortex and superior

temporal gyrus were higher compared to wakefulness (Figure 2.7b). In the

comparison between wakefulness and SWS (Figure 2.7c), functional connectivity of

the bilateral IPL, PCC and mPFC with the HF was stronger in wakefulness, indicating

a sleep stage dependent decrease in the integration of the HF to the DMN. No

regions exhibited stronger functional connectivity with the HF during SWS than

wakefulness. The comparison between S2 and SWS (Figure 2.7d) revealed stronger

HF connectivity to occipital, lateral temporal and inferior frontal regions in S2. No

clusters were detected that showed higher functional connectivity during SWS

compared with S2.


Figure 2.7. Comparison of HF functional connectivity across sleep stages. Wakefulness: W;

sleep stages 1 and 2: S1/S2; slow wave sleep: SWS. Data are thresholded at puncorr<0.001, with

a variable voxel extent, resulting in pFWE,cluster<0.05 (see Table A2.2, appendix A2). Color coding

indicates t-values. The HF has been masked out to avoid display of autocorrelations. MNI

coordinates of each slice are given.

Subregions analysis

We further tested specific functional connectivity of three HF subregions

across sleep stages. An F-test on the main effect of subregion revealed wide-spread

differences in functional connectivity among subregions (Figure 2.8).

Figure 2.8. Results of the full-factorial design employing factor sleep stage and subregion. F-

test showing main effects of subregion. Data are thresholded at pFWE,cluster<0.05 using a


collection threshold of p<0.001. Note that the HF has not been masked out. MNI coordinates of

each slice are given. Color bars depict t-values.

Post-hoc t-tests (Figures 2.9 and Table A2.3, Appendix A2) showed that

during wakefulness, functional connectivity of the CA was most robust to the IPL, the

midline nodes of the DMN, and the thalamus. The functional connectivity with the

lateral temporal cortex appears to be strongest for the CA in wakefulness and all

NREM sleep stages. In S1, CA showed unique functional connectivity to the primary

motor cortex. In sleep stage S2, SUB revealed strongest functional connectivity to the

anterior cingulate cortex (ACC)/mPFC. Finally, in SWS, the DG showed unique

functional connectivity to the occipital cortex, while SUB was most connected with the

motor cortex.

Figure 2.9. Comparison of subregional HF functional connectivity contribution within sleep

stages. For each sleep stage, strongest contributions of each hippocampal subregion are

depicted (yellow: dentate gyrus (DG), red: cornu ammonis (CA), green: subiculum (SUB)). (a)


Wakefulness (W); (b) sleep stage 1 (S1); (c) sleep stage 2 (S2); (d) slow wave sleep (SWS). Data

are thresholded at pFWE,cluster<0.05 using a collection threshold of p<0.001 (see Table A2.3,

appendix A2). The HF has been masked out to avoid display of autocorrelations. MNI


Sleep spindles

Further, Activation maps related to sleep spindle activity in sleep stage 2,

differentiating between slow and fast spindles were generated. The basic pattern

reported by Schabus and colleagues [64] for fast spindles, with major clusters in the

right thalamus, insula, lateral temporal cortex, cingulate cortex and motor areas could

be reproduced (Figure 2.10 and Table A2.4, Appendix A2). Notably, no spindle

specific activation was found in the HF, in accordance with the previous report.

However, we observed a strong overlap between the HF functional connectivity map

(contrast S2>W; Figure 2.7) and the network associated with fast spindles. Neither

significant positive or negative correlations with slow spindles, nor negative

correlations with fast spindles were observed.

Figure 2.10. Activity related to fast sleep spindles in S2 (pFWE,cluster < 0.05, collection threshold

p<0.001). Color bar depicts T-values. Note that the HF is not part of the spindle network. MNI


In order to investigate if observed increases in hippocampal functional

connectivity during S2 compared with wakefulness are associated with spindle

activity in the EEG data, PPI analyses separately for the three hippocampal


subregions within S2 were performed. Functional connectivity of the subiculum

increased in parallel with EEG spindle activity (Figure 2.11, Table A2.5, Appendix

A2) in a network comprising the cingulate cortex, lateral temporal cortex, motor

cortex, SMA, and insular cortex. Spatially similar, yet less strong spindle-associated

functional connectivity increases were found for the CA and DG bound network

(Table A2.6 and A2.7, Appendix A2).

Figure 2.11. PPI analysis for the interaction spindles x SUB (hot colors, puncorr<10-4, extent>150

voxel). Color coding indicates t-values. MNI coordinates of each slice are given.

Chapter 2 66

Discussion

The present data demonstrate altered functional connectivity of the

hippocampal formation across NREM sleep stages compared with the waking state.

The previously reported integration of the hippocampal formation into the DMN during

wakefulness was confirmed. During sleep stage 2, the HF was more strongly

connected with the temporal, insular, occipital and cingulate cortices than during

wakefulness or slow wave sleep. Spindle related activity overlapped with the HF

connectivity map (contrasting sleep stage 2 with wakefulness). Especially the

connectivity pattern of the subiculum to frontal, lateral temporal, and motor cortical

regions and to the insula showed a strong interaction with occurrence of sleep

spindles.

HF connectivity to the DMN in wakefulness and sleep

Contribution of the HF to the DMN has been reported earlier [104;105].

Possible functions of the DMN include internal mentation and autobiographic memory

retrieval: In a meta-analysis of autobiographical memory tasks, the HF, mPFC and

RspC have been found to be robustly co-activated [115]. Similarly, episodic memory

retrieval is related to activation in the lateral posterior parietal cortex and the

precuneus extending into PCC/RspC [116]. Burianova et al. [117] showed that a

common neural network including major DMN nodes underlies the retrieval of

declarative memories during wakefulness, independent of specific memory content.

Recently, Andrews-Hanna et al. [118] showed that the DMN is comprised of two

interacting subsystems, linked by a common midline core (PCC; mPFC). One of

these subsystems, the medial temporal lobe subsystem, which includes the HF and

Chapter 2 - Discussion 67

the IPL, is particularly engaged during construction of mental scenes based on

memory. Taken together, these reports showed that the functional connectivity of the

hippocampus and the DMN conjointly constitute a network pivotal for episodic

memory retrieval during wakefulness.

Only a trend for reduced coupling of the HF to the DMN in light sleep stages

was observed. In contrast, slow wave sleep was characterized by a robust decrease

in functional coupling between the HF and the DMN. This breakdown of functional

connectivity between the HF and DMN occurred in all core nodes of the DMN (PCC,

mPFC, bilateral IPL). Reduced connectivity between the anterior and the posterior

DMN nodes in SWS have been reported before [107;108], although the posterior

nodes (PCC, IPL) seemed to increase their connectivity [107]. Here, these findings

were extended by demonstrating that the breakdown of connectivity within the DMN

also comprises the HF. Reduced functional connectivity of the hippocampus to core

DMN nodes (PCC, IPL) has also been reported for subjects under anaesthesia [119],

which may point towards a common neural basis for behavioural similarities during

these stages.

HF connectivity to neocortex generally higher in S2 than in SW

The present data showed that functional connectivity of HF with neocortical

regions was higher in S2 than in SWS or wakefulness. This indicates functionally

distinct processes during sleep: increased connectivity between HF and neocortical

regions in S2 suggests an increased capacity for possible information transfer,

whereas decreased connectivity between these regions in SWS suggests a

functional system optimal for segregated reprocessing. These findings may be

relevant for the memory consolidation hypothesis of sleep [75;76], and may inform

during which NREM sleep stages memory transfer and reprocessing could occur.


Several studies stress the relevance of SWS for memory consolidation

[73;78;85;120;121]. However, there are reports that S2 may in fact play an equally

important role [86;93;122-124]. Furthermore, some data on the relevance of sleep

EEG macro-structure for memory consolidation are derived from animal models, in

which NREM sleep is usually not divided into substages [125]. Hence it is not

unambiguously clear, how, within NREM sleep, S2 and SWS differentially relate to

the hippocampal-neocortical dialogue proposed to underlie sleep-related memory

consolidation. Functional connectivity analyses may help to dissect such sleep stage

specific contributions.

The analysis of subregional contributions to the HF network revealed that the

CA region is driving the functional connectivity with the core DMN nodes (PCC/RspC,

IPL, mPFC) during wakefulness and to the lateral temporal cortex in all sleep stages.

In contrast, mPFC functional connectivity is dominated by the subiculum in S2. In a

series of elegant fMRI studies, sleep has been shown to alter the hippocampal-

neocortical interplay, e.g. with the mPFC, underlying memory retrieval [92;126-131].

The classical model of memory consolidation suggests that information is

propagating from the neocortex to the hippocampus in the waking state while being

reversed in sleep, e.g. [90], however, a recent study showed more bidirectional

coupling with increased influence of neocortical regions onto the HF towards deeper

sleep stages [132]. The present data don’t allow commenting on the direction of such

information propagation due to intrinsic limitations of functional connectivity analysis.

Still, the strong involvement of the subiculum as the major hippocampal output region

in S2 may suggest a hippocampal-to-neocortical information transfer.


Increased HF connectivity in S2 interacts with sleep spindles

The data reveal that the connectivity between HF and neocortex showed a

strong interaction with fast sleep spindles, which was most pronounced for functional

connectivity between the subiculum and the lateral temporal, insular, cingulate and

medial prefrontal cortices. Interestingly there was no significant activation of the HF in

direct association with occurrence of either fast or slow spindles (Figure 2.10) in

accordance with Schabus et al. [64]. This suggests that spindle activity may enhance

functional connectivity between the HF and neocortical regions, but that it is not the

sole cause of such connectivity. Temporal coupling between spindles, hippocampal

ripples and slow-oscillations has been described before in that during the up-state of

cortical slow oscillations, both spindle activity and hippocampal high-frequency ripple

activity were increased [133;134]. The coordinated spindle–ripple events have been

suggested to provide a mechanism for information transfer between hippocampus

and neocortex [94;133;135], which could explain that sleep spindles, most

pronounced in S2 but also occurring in SWS [112] are related to memory

consolidation [86;122;123;136-138]. At the same time, sleep spindles have been

shown to reflect thalamus-driven cortical inhibition, which may signify a different or

double functionality [65;66].

Connectivity between the thalamus and HF was reduced in S1. This thalamic

cluster was located in the nonspecific midline nucleus, known to send input to the

hippocampal formation [139;140]. The finding of reduced HF/thalamus functional

connectivity is in line with earlier observations of thalamic deactivation at sleep onset

[10] and with generally reduced thalamic functional connectivity to most cortical brain

regions in S1 [114].


Limitations

While the present data provided evidence for generally increased HF

functional connectivity in S2 and a relation of the hippocampal functional network and

sleep spindles, one cannot conclusively state a direct link to memory related

processes. Future studies will need to compare HF functional connectivity in S2 with

successful post-sleep memory recall after intervention, as has been demonstrated in

wakefulness [88;92]. Alternative functional interpretations, such as synaptic

downscaling [70;141] should also be considered in future experiments.

Conclusion

The present analysis revealed important reorganization of the spontaneous

HF connectivity during NREM sleep. Especially during sleep stage 2 and linked to

sleep spindle activity, synchronous activation with lateral temporal, cingulate and

frontal regions arises which may represent coordinated neural activity within the

memory network. As fMRI functional connectivity allows tracking of memory transfer

in wakefulness, these altered functional connectivity patterns likely reflect differences

in information processing, and possibly in sleep specific plasticity processes.

71

CHAPTER 3

Recurrence quantification analysis allows for single-trial stimulus

differentiation in evoked potential studies

Introduction

Event related potential (ERP) studies investigate specific deflections in

electroencephalographic recordings associated with a given stimulus. For example,

an auditory oddball experiment consists of a sequential presentation of tones in

which frequent tones (presented at one pitch) are interspersed with rarely occurring

‘odd’ tones (also referred to as rare tones, presented at a different, e. g. a higher

pitch) (Figure 3.1). This experiment typically elicits the ERP components N100, P200

and P300. N (negative) or P (positive) describes the observed polarity of the evoked

signal, and the number the typical time of its occurrence after the stimulus onset. The

P300 deflection represents aspects of cognitive processing such as attention, novelty

detection and memory updating and has been extensively described in the literature

[142-144]. The P300 amplitude as obtained from averaged individual ERPs is

inversely proportional to the frequency of the tones’ appearance [142;145] which

allows for differentiation of the neural response to frequent and rare tones. The N100

and P200 deflections are related to perceptual processing and are dependent on the

subject’s attention level [146;147]. N100/P200 are further reflecting physical tone

characteristics like pitch and loudness.

ERPs are usually lower in amplitude compared with the EEG background

signal. Therefore, ERPs are usually averaged over several trials to reduce the

contribution of random EEG background activity, and to increase the signal-to-noise

ratio.


Figure 3.1. Active auditory oddball experiment. (a) Frequent tones (black) intercalated by rare

(red) tones with inter-stimulus interval (ISI) of 1000 ms. The subject was asked to press a

response button every time he detected a rare tone. (b) ERP response to rare tone displaying

N100, P200 and P300 peaks. (c) Topographic map from the blue time window in (b).

Such averaging intrinsically assumes that the electrophysiological responses

are similar for all stimuli with the same physical characteristics, i.e. that all ERPs are

stationary. This assumption, however, does not necessarily hold true for physiological

data. Individual responses to stimuli depend on levels of attention, habituation,

vigilance and other factors [145;148-151] and averaging does not account for such

potentially meaningful trial-by-trial variance. Furthermore, averaging methods

collapse single events which are gathered over an extended period of time (several

minutes). This may affect, for example, simultaneous ERP/fMRI analyses, in which

correlation of each single ERP with its corresponding fMRI volume is required.


Therefore, particularly for event-related fMRI analysis, methods are sought that allow

for a robust quantification of the electrophysiological response to a single trial.

Several research groups have reported advanced methods to detect minute

changes in the ERP of a single trial. Bénar et al. [152] were able to track the P300

response changes in an EEG/fMRI oddball experiment applying narrow filters in the

EEG data. Similarly, independent component analysis (ICA) and wavelet denoising

have been proposed to improve trial-by-trial analysis [153-157]. However, all these

approaches have their specific advantages and limitations, and no standard method

for single trial analysis has yet been established [158].

Recurrence quantification analysis (RQA) is a method able to detect linear

and nonlinear signal changes proposed by Zbilut and Webber [159]. It does not

require any specific assumptions about the statistical properties of the data. In

contrast to other non-linear methods, RQA can also be applied to relatively short non-

stationary time series [160]. Carrubba et al. [161-163] compared the classical method

of time averaging and recurrence analysis to detect magnetic evoked potential. They

reported that these potentials were better detected by RQA than by the averaging

method. Schinkel et al. [164] investigated ERP response to semantic mismatch,

known to evoke a N400 deflection: On a trial-by-trial level, they compared the

sensitivity of the ERP amplitude analysis with RQA and showed that RQA was more

sensitive in detecting semantic mismatch. Marwan and Meinke [165] demonstrated

that RQA could also be used to analyze single-trial ERPs in an auditory oddball

paradigm, however, no group analyses were presented.

The MR induced artifacts in the EEG, caused by the static magnetic field and

by rapidly switching magnetic field gradient, make the potential combination of single-


trial EEG and fMRI even more challenging (Chapter 1). Until now no work was done

to probe RQA in simultaneous EEG/fMRI.

Objective

The goal of this study was to test the reliability of RQA to characterize single

trial ERPs recorded under normal experimental conditions as well as under the

influence of MR artifacts, and to compare the results with ERP amplitude analysis.

Furthermore, we set out to extend previous reports from an individual subject level to

the level of group analysis, allowing for statistical testing.

Approach

A classical auditory oddball experiment was used to compare the

performance of RQA with the performance of amplitude analysis of ERPs.

Specifically, RQA was compared to standard ERP amplitude analysis with respect to

their potential to distinguish between frequent and rare tones. Second, these

methods were compared with respect to their power to distinguish between frequent

tones positioned before and after a rare tone, as the appearance of a rare tone may

change the context of the experimental background. Last, for a preparation of

combined fMRI/EEG analysis, the robustness of RQA and amplitude analysis

towards experimental noise in the raw data was investigated. For this purpose, EEG

data were recorded under standard laboratory conditions as well as during MRI

acquisition.

Chapter 3 75

Methods

Subjects

Eleven right-handed healthy subjects (7 men, 4 women; mean age 27.8 ± 2.8

years) were recruited by public advertisements. They underwent a diagnostic

interview to exclude any neurological, psychiatric or medical condition which could

prohibit participation in the study, and to exclude any contraindication for MRI. The

study protocol followed the guidelines of the Declaration of Helsinki and was

approved by the local ethical committee. All subjects gave their written informed

consent and were paid for their participation. Data of two subjects were excluded due

to strong movement artifacts, leaving nine subjects for final analysis.

Experimental task

Subjects performed two sessions of an active auditory oddball paradigm [166-

168], outside and inside the MR scanner, on the same day. Presentation software

(Neurobehavioral Systems, Albany, USA) was used for task programming and

stimulus delivery. Each section consisted of the presentation of two types of tones: A

rare tone (1.4 kHz, duration 100 ms including 10 ms rise and fall times) and a

frequent tone (1 kHz, duration 100 ms including 10 ms rise and fall times) that may

be considered the auditory background pattern. The order of the tones was randomly

assigned, with two consecutive rare tones always being separated by at least two

frequent tones. The overall frequency of appearance was 90% for frequent and 10%

for rare tones. This distribution criterion had to be fulfilled for all subsets of 20 tones.

The inter-stimulus interval (ISI) was 1 s with jittering of ± 250 ms. For each section a


total of 600 tones were presented over 15 minutes. Tone stimuli were presented via

headphones, and subjects wore ear plugs as a safety requirement during MR

measurements. All subjects participated in an MR test session to become acquainted

to the MR environment prior to the actual experiment. To adjust loudness of the

auditory stimuli, a preparation scan was performed during which subjects had to

repeatedly decide whether or not they perceived the tones as loud as the fMRI

scanner sound. This resulted in a defined level of subjectively identical loudness of

tones and scanner noise. For the final experiment, tones were delivered 3 dB louder

for better perception (within an absolute range of 80 – 85 dB). Subjects were

instructed to keep the eyes open in order to reduce EEG alpha activity, and to press

a button with their right index finger as fast and accurately as possible every time

they detected a rare tone.

MRI data acquisition

In the simultaneous EEG/MRI experiments, each subject was positioned

inside the MR scanner in a supine position with the head carefully immobilized to

minimize movement artifacts. Experiments were performed on a 1.5 Tesla clinical

scanner (Signa Echospeed, General Electric, Milwaukee, Wisconsin, USA) using an

8-channel headcoil. Functional T2*-weighted echoplanar images (EPI) of the whole

head (342 volumes) were obtained with 25 slices and an image matrix size 64 × 64

(single shot EPI, TR = 2 s, TE = 40 ms, resolution of 3.44 × 3.44 × 4 mm3, FOV = 22

× 22 cm2). Slice orientation was parallel to the anterior/posterior commissure.


EEG acquisition and preprocessing

EEG acquisition was performed at a sampling rate of 5 kHz using an MR-

compatible system (Brain Products, EasyCap, Herrsching-Breitbrunn, Germany) with

19 electrodes positioned according to the 10-20 system referenced against FCz. In

addition, EMG, EOG and ECG were acquired and used for artifact correction. All

precautions to guarantee a safe recording of the electrophysiological signals during

image acquisition were taken [39]. EEG preprocessing was performed using Brain

Vision Analyzer software 1.05 (Brain Products, Gilching, Germany). Correction of

MRI related EEG artifact was performed by subtraction of an adaptive template

based on artifact averages of eight consecutive MRI volumes. Independent

component analysis was applied for removal of ECG and EOG artifacts. Components

with strong ECG and EOG related time courses were visually identified and removed.

Data were resampled to 250 Hz, re-referenced against linked mastoids and filtered

(0.1 Hz to 30 Hz) (see Chapter 1 for more details about EEG preprocessing). Similar

to Benár et al. [152], an additional filter to decrease the power of the EEG alpha band

was applied. For this purpose, a high frequency cut-off of 10 Hz was used, mainly

reducing alpha components in the range 10-12 Hz, but retaining the lower frequency

range containing the individual ERPs. Finally, the EEG was segmented according to

each tone stimulus, using a time window of 200 ms prior to tone onset to 800 ms

after tone onset (-200 ms to +800 ms), DC detrended and corrected for different

baseline levels by subtracting the average amplitude between -200 ms and 0 ms. All

further analyses were restricted to the three central EEG derivations Fz, Cz and Pz

as most commonly selected for ERP studies [142;145;166;169;170].


Recurrence Quantification Analysis

The state of a system, e.g. atmospheric, biologic and physiologic systems,

can change in time. Due to the importance of such alterations for human life, several

research groups focus their effort on finding ways to describe and, of course, predict

the dynamics of these fluctuations.

The numbers of variable needed to describe a system depends on how

complex it is (each variable represents one dimension of its phase space). When real

systems are analyzed, usually not all variables (and states) necessary to fully

describe it can be measured. Takens’s theorem [171] demonstrated that it is possible

to reconstruct a picture that contains the same topological information of the

multidimensional system behaviour using only one variable. In this theorem, the

multi-dimensional state vector xr

that describes the system trajectory in an m-

dimensional space can be constructed from a one-dimensional time series iu :

),,...,,( )1( dmidiii uuux −++=r

(3.1)

where d is the time delay and m is the embedding dimension. m should be greater

than two times the original dimension of the system to the theorem be validated.

In 1987, Eckmann et al. [172] proposed a visual tool called recurrence plot

(RP) to graphically depict temporally correlated linear and non-linear information of

time series. The RP construction uses Equation (3.1) to create the states of the

system, and visualizes those states that recur in time. The RP can be graphically

represented by a binary NN × matrix with black and white dots representing two

states being close or distant to each other at a given threshold (ε ). Thus, the RP

matrix is defined as:

( ).R , jiji xxrr

−−Θ= ε (3.2)


)(yΘ is the Heaviside function: 0)( =Θ y (white dot in the RP) if 0<y , and 1)( =Θ y

otherwise (black dot in RP); Nji ,,1, K= , with N being the number of states and

being a matrix norm that measures the distance between two vectors. In the present

work, the Euclidian norm was used. Therefore, the dynamic information of the system

is eventually contained in the structure of the RP. As an example, a random system

shows just isolated black dots in the RP, as system recurrence is random. A sinusoid

fluctuation, on the other hand, shows periodic diagonal lines in the RP, the distance

of which relates to the basic frequency of the system (Figure 3.2).

Figure 3.2. Time series (top row) and recurrence plot (bottom row) of a random (left) system

and a sinusoid (right). Settings were: 4=m , 3=d and 10 % of fixed amount of nearest

neighbors. The main diagonal in both RP represents the degree of auto-recurrence, i.e. the

comparison of each state with itself.

The choice of the distance threshold ε is crucial. If ε is too small, few

recurrence points will be found and information about the dynamic of the system will

be lost. If ε is too high, many neighbours will be falsely classified as recurrent,

resulting in a lot of artifacts in the RP. In the present study, each state vector xr

has


the same percentage of neighbours (10 %). The ixr

neighbours are defined as jxr

contained in a sphere centred in ixr

(fixed amount of nearest neighbours- FAN). By

this a different ε is attributed to each point to achieve a constant recurrence density

[173].

RP is a useful method to visualize dynamic behaviour, but depending on the

complexity of the system, additional quantification of the RP is needed. For this

reason, Zbilut and Webber [159] introduced a quantification scheme of the

recurrence points in a RP as a whole (percentage of the RP showing recurrent

points) and of specific segment lines in the RP (diagonal lines: 1R , =++ kjki , lk K,1=

and l is the diagonal line length). Marwan and co-workers introduced measures

based on vertical lines ( 1R , =+kji , vk K,1= , v is the vertical line length) and showed

that these measures are capable to detect chaos-chaos transitions in physiological

data [174].

In the present work, RQA was performed using the Matlab (R2008b version,

the MathWorks, Inc., USA) and the CRP toolbox version 5.5 (http://www.agnld.uni-

potsdam.de/~marwan/toolbox). The RQA variables analyzed in this study are defined

in Table 3.1, matching also the CRP toolbox manual version 5.15, release 28.

The choice of the embedding parameters dimension and delay is also

critically influencing the RQA results. Common methods for their identification are the

false nearest neighbours method for the dimension (m ), as proposed by Kennel et al.

[175], and the mutual information approach for the delay (d ) as suggested by Frazer

and Swinney [176] for non-biological time series. However, it is still unclear if these

approaches are also valid for biological systems. Consequently, RQA using different

combinations of m and d was applied to screen how the power of RQA variables to

discriminate tone types depends on these parameters (Figure 3.3).


3.1. RQA variables

Variables based on diagonal lines Variables based on vertical lines

Determinism (DET)

∑

∑==N

ji

m

ji

N

ll

R

llP

,

,,

)(

DET min

ε

ε

li NillP ...1,)( ==ε is the

frequency distribution of diagonal

structures with lengths l , lN in the

total number of diagonal lines and

2min =l the defined minimum

length of the diagonal structure.

Laminarity (LAM)

)(

)(

LAM

1

min

vvP

vvP

N

v

N

vv

∑

∑

=

==

ε

ε

vi NivvP ...1,)( ==ε is the

frequency distribution of vertical

structures with lengths v , vN in

the total number of diagonal lines

and 2min =v , the defined

minimum length of the vertical

structure.

Maximal diagonal

line (MaxL)

( )li NilMaxL K1;max ==

Trapping time (TT)

)(

)(

TT

min

min

vP

vvP

N

v

N

vv

∑

∑

=

==ε

ε

Average diagonal

line (AvgL) ∑

∑

=

==

l

l

N

ll

N

ll

lP

llP

min

min

)(

)(

AvgLε

ε

Maximal length of

vertical structures

(MaxV)

( )vi NivMaxV K1;max ==

Table 3.1. RQA variables definition as described in the manual of the CRP toolbox, Version

5.15, Release 28.

The number of state vectors created from a time series is given by

,)1( dmw −− with w being window size, i.e. e number of points used from the one-

dimensional time series. In the present work, the window size was 100 points, as

described below. m and d combinations were limited to values that produced at

least 50 states (Figure 3.3).

The respective range of the RQA variable maximum diagonal line (MaxL),

average diagonal line (AvgL), trapping time (TT) and maximum vertical line (MaxV)


depends on the number of the analysed states. In order to be able to compare RQA

values obtained with different m and d , the original values were divided by

.)1( dmw −−

Definition of ERPs and RQA time windows

The P200 and P300 were defined as maximum EEG amplitude between 160

- 252 ms and 260 - 460 ms, respectively [144;145;148;166]. N100 was defined as the

minimum peak in the time window 50 - 156 ms [153].

A topographic map of the grand average of the ERP of the difference

between rare and frequent tones (time window: 0 ms to 800 ms after tone

presentation) was generated using the triangulation and linear interpolation

algorithms as implemented in the Brain Vision Analyzer version 1.5. A topographic

map of the ERP difference between frequent tones immediately preceding rare tones

(frequentbefore) and frequent tones immediately following rare tones (frequentafter) was

also generated.

While the analysis of ERP components using amplitude averages is rather

standardized, it was a priori unknown which time window would deliver the most

sensitive RQA measurements. The window size was chosen based upon a tradeoff

between temporal resolution and reliability of the estimate. A large window results in

good estimates but low resolution, while a small window results in good resolution but

poorer estimates. Therefore, three different time windows for RQA each of which

comprised 100 sample points (W1: 84 – 484 ms, W2: 204 – 604 ms, and W3: 324 –

724 ms) were chosen for analysis. These broader time windows cover a total of

about 800 ms after tones and allow testing of high values of m and d and still having

good estimates.


Each ERP amplitude and RQA variable was identified for all trials as

described above and then compared between rare and frequent tones (Figure 3.3).

To control for any bias in the analysis possibly introduced by the discrepancy

between the specific temporal windows for ERP amplitude detection and the ones

chosen for RQA, the mean amplitude and the mean absolute amplitude of the full

temporal windows W1, W2, W3 were also analyzed for their potential to differentiate

between tone types.

Statistical analysis

First, the six RQA variables (Table 3.1) and the nine amplitude based ones

were explored, including ERP, mean and mean absolute amplitude (amplitude

analysis - AA) with regard to their potential to discriminate between frequent and rare

tones and focusing on data recorded without concurrent MRI acquisition. As the

resulting RQA and AA measures cover different absolute ranges, these measures

were subjected to a z transformation ( σµ)( −= xz ); x , µ and σ are the raw value,

mean and standard deviation of all trials, respectively) to allow for comparison across

different measures. Transformed measures were then used to calculate Cohen's d as

effect size measure for the difference between rare and frequent tones for each

subject. Statistical inference at the group level was then estimated by one sample t-

tests against zero for the respective measures. This allowed to control if effects

shown the same sign across subjects. For each of the two methods (RQA versus

AA), the variables with highest mean effect size (see results) was used for the further

analyses. To compare RQA with AA, a paired t-test between the two optimal

measures was performed (Figure 3.3).


Figure 3.3. RQA and AA pipeline. (a) RQA procedure: (a.1) for each single trial three time

windows were chosen to apply RQA (W1: 84 – 484 ms, W2: 204 – 604 ms, and W3: 324 – 724

ms). (a.2) RPs are created and the RQA variables are quantified (Table 3.1). (a.3) the procedure

in (a.2) is repeated for several values of m and d . (a.4) All values found in procedure (a.3) are

summed following the present equation (RQAsum). All the previous steps are repeated for each

trial (rare and frequent tones). Step (a.5) exemplifies a histogram graph of all trials for one

RQAsum variable. (a.6) These RQAsum values are z-transformed. (a.7) the trials are grouped in

rare and frequent tones. (a.8) Cohen effect sizes are calculated. RQA

rareµ and RQA

frequentµ are the

mean of the RQAsum values for all rare and frequent tones respectively.

( ) ( )

+−= )(1

2

frequentrare

RQA

ii

RQA

i nns

ns , where RQA

is is the SD of the RQAsum values, i = rare or

frequent and raren and frequentn are the total number of rare and frequent trials, respectively.

Steps (a.1) to (a.8) are repeated for each subject and EEG channels (Fz, Cz, Pz). (b) AA

analysis: (b.9) P300, P200, N100, mean and squared mean amplitudes (W1, W2, W3) are

indentified for each trail. Note that for reporting ERPs, positive deflections are showing

downward maxima, while negative deflections are upward. (b.10) exemplifies a histogram

graph of all trials for one AA variable. (b.11) These AA values are z-transformed. (b.12) the

trials are grouped in rare and frequent tones. (b.13) Cohen effect sizes are calculated. AA

rareµ and


AA

frequentµ are the mean of the AA values for all rare and frequent tones respectively.

( ) ( )

+−= )(1

2

frequentrare

AA

ii

AA

i nns

ns , where AA

is is the SD of the AA values, i = rare or

frequent and raren and frequentn are the total number of rare and frequent trials respectively

Steps (b.9) to (b.13) are repeated for each subject and EEG channels (Fz, Cz, Pz). (c) For each

method (RQA and AA), the variable with higher mean Cohen effect size between subjects was

chosen as optimal variables. The methods were cross-tested by a paired t-test between the

Cohen effect sizes of these optimal variables (Optimal

RQACohen and Optimal

AACohen for RQA and AA

respectively).

Additionally, it was assumed that both amplitude and RQA measures show a

high degree of temporal auto-correlation over the course of an experiment. This auto-

correlation is in part caused by subject-specific features such as head geometry and

skin conductance properties that lead to higher within-subject similarity of trials

compared with between-subject similarity. Second, it is caused by slow amplitude

fluctuations or slow changes of non-linear patterns that - across several trials – may

be induced by fluctuations of the brain’s general functional state, e.g. reflecting

alterations in vigilance or attention [145;148-151]. To account for this auto-correlation

over the experiment and to optimize the sensitivity of both methods, a second

analysis was performed in which frequent or rare tones were not pooled all together.

Instead, pairs of rare tones and the preceding or following frequent tone were

grouped together. For the ‘optimal’ variables identified in the previous step (Figure

3.3), the difference of these consecutive tones per trial was calculated and difference

values gained for RQA (∆RQA) and the amplitude analysis (∆AA). The temporal order to

calculate the difference values between tones was 'previous minus the following trial'.

For both methods, these difference values were entered into one-factorial repeated

measures ANOVA (factor trial with T levels, where T is the number of trials). Here,

the significant changes over the experiment (trial factor) were not in the focus, but the


significance of the deviation of all differences from zero represented by the intercept.

Importantly, this approach allowed for a trial-wise comparison across methods by

performing an across method subtraction (∆RQA-∆AA) and testing the resulting

difference values against zero in a repeated measures ANOVA, again testing the

intercept. If the intercept factor had a significant effect and the estimated means of

the factor trial showed to be positive or negative, this would represent a superiority of

RQA or AA method, respectively.

Chapter 3 87

Results

Explorative analysis of ERP amplitude and RQA variables

First, the RQA (Table 3.1) and AA variables (N100, P200, P300, mean

amplitudes and absolute mean amplitudes of time windows W1, W2, W3) derived

from the central electrodes (Fz, Cz and Pz) were investigated. This investigation was

based on the dataset obtained outside the MR environment, focussing on Cohen's d

as effect size measure in the comparison of frequent and rare tones.

Figure 3.5 shows the grand average of the ERPs in response to frequent and

rare tones (left column), and the spatial representation of the difference between both

time series as topographical map (mid column, left image). As expected, P300

amplitudes were higher (more positive deflection) for rare tones. The highest

amplitude difference between rare and frequent ERP grand averages was found in

posterior brain areas.

For the AA measures, highest mean effect size value was found for the P300

time window (0.90 ± 0.47) in Pz channel (Figure 3.4a).

As mentioned, RQA variables are highly dependent on the embedding

parameters m and d . This dependency showed a huge inter- and intra-subject

variability (see Figure 3.6). Thus, to avoid a suboptimal choice of m and d , a sum

score based on all combinations of 1 ≤ m ≤ 20 and 1 ≤ d ≤ 20 was used (Figure 3.3,

(a.3)).


Figure 3.4 Differentiation power of AA and RQA group analysis for undistorted EEG recordings.

(a) Effect sizes of the group mean amplitude and 95% confidence interval for N100, P200, P300

(bars 1-3) as well as mean and mean absolute amplitudes (bars 4-9), results are depicted

separately for electrodes Fz, Cz and Pz. Note the nominally highest average effect for the P300

amplitude in channel Pz. * indicates not significant group effect (p>0.05). (b) Effect sizes of

RQA variables for all midline electrodes and time windows W1, W2, W3. Superiority of AvgL in

time windows W1 for all electrode positions can be seen. (c) Strong cross-correlation was

noted between the six selected RQA variables (r>0.377, p<0.01).


Figure 3.5. Topography of ERP grand averages (data acquired outside the MR). Left column:

Grand averages of rare tones (1), frequent tones (2) and the difference between them (1-2).

Right column: Grand averages of frequentbefore(3), frequentafter (4) and the difference between

them (3-4). Middle column: EEG topographic mapping of (1-2) and (3-4).

This analysis approach prevents overfitting of the specific experiment by

selecting a fixed set of embedding parameters. The alternative approach - optimizing

m and d separately for each subject – may hamper interindividual comparison, as

e.g. in group studies.


Figure 3.6. Cohen’s effect sizes for comparison of frequent and rare tones using RQA measure

AvgL. Effect sizes for different embedding parameters m and d were calculated for each

subject on the basis of undistorted EEG data (Pz electrode) and are depicted as color-code.

While considerable inter-subject variance of the effect sizes and their distribution across m

and d were seen, the range of m between about 2 and 5 provided effect sizes >0.5 across

subjects.

Using the sum scores (Figure 3.3, (a.4)), AvgLsum measured in W1 of Pz

showed highest mean effect sizes (0.91 ± 0.22) for separation of rare and frequent

tones (Figure 3.4). Therefore, the electrode Pz and time window W1 was used for

further analyses.

RQA variables are highly correlated among each other (Figure. 3.4c),

suggesting that the results of different values of m and d may have validity also for

RQA variables other than AvgL. As mentioned above, it was not possible to identify a


single m and d pair which would allow for optimized tone differentiation in all

subjects. However, for AvgL highest effect sizes for tone discrimination were

generally obtained for 2 ≤ m ≤ 5 and 5 ≤ d ≤ 20. Given that the comparison between

AvgLsum calculated with 1 ≤ m ≤ 20 and 1 ≤ d ≤ 2 or 2 ≤ m ≤ 5 and 5 ≤ d ≤ 20

showed significant results (t(8) = 5.04, p = 0.001, paired t-test), further steps of the

analyses were restricted to 2 ≤ m ≤ 5 and 5 ≤ d ≤ 20 and the reported RQA results

were based on the following variable and embedding parameters (Figure 3.6):

50))1((|,,))1((

20

5

5

2

, ≥−−Ζ∈−−

=∑∑= =

dmwmddmw

AvgLAvgL

d m

md

sum (3.3)

To compare between the classical ERP amplitude analysis and the RQA,

P300 and Equation (3.3) for time window W1, both measured in Pz were used. The

power to distinguish between rare and frequent tones for data recorded with and

without concurrent MRI acquisition was estimated.

Cross-methods comparison of differentiation between rare and frequent tones

(EEG recordings outside the scanner)

At the group level, both P300 (0.90 ± 0.47, t(8) = 5.79, p <0.0001) and

AvgLsum (0.91 ± 0.22, t(8) = 12.38, p < 0.0001) showed similarly high effect sizes with

lower variance across subjects seen for AvgLsum. Effect size of AvgLsum were higher

than those of P300 in six out of nine subjects (Table 3.2), yet, formal pair-wise

comparison between the methods yielded no significant difference (paired t-test, t(8)

= 0.58, p = 0.58).


3.2. Effect sizes for AvgL and P300 gained from undistorted EEG recordings

Subject AvgLsum P300

1 1.01 0.79

2 1.08 0.47

3 0.84 0.62

4 1.25 0.47

5 1.09 1.95

6 0.80 0.65

7 0.70 1.22

8 0.89 0.10

9 0.54 0.94

Table 3.2. Cohen’s effect size (rare vs. frequent tones) for P300 (channel Pz) and AvgLsum

(channel Pz; time window W1) for each individual subject.

Cross-methods comparison of differentiation between subtypes of frequent

tones (EEG recordings outside the scanner)

As stated above, the occurrence of a rare tone interrupting the continuous

series of frequent tones may lead to changes in neural responsiveness, e.g. by

increasing the subject’s arousal. Therefore, differences across the AA and RQA

methods were calculated for corresponding frequent trials flanking a rare tone.

Statistical inference was calculated comparing all such pairwise differences against

zero by using repeated measures ANOVA. Of a typical series of trials (frequentbefore -

rare - frequentafter), i) rare tones with respect to the preceding frequent tone

(frequentbefore/rare), ii) rare tones with respect to the following frequent tone


(rare/frequentafter), iii) frequent tones flanking a rare tone (frequentbefore/frequentafter)

were compared.

For both P300 and AvgLsum strong differences for tone pairs (i) and (ii) were

detected (Table 3.3), yet, both analytical methods were not significantly different from

each other (F(1,8) = 0.99, p = 0.350, and F(1,8) = 1.08, p = 0.330 for (i) and (ii),

respectively). In contrary, only AvgLsum but not P300 was effective in distinguishing

between frequentbefore and frequentafter tones (pair (iii)) (Table 3b).

The ERP grand averages of the difference between frequentbefore and

frequentafter showed a different topographic distribution (Figure 3.5, right column)

compared with the frequent/rare comparison (Fig 3.5, left column) with similar

absolute amplitudes in all midline electrode positions (Fz, Cz, Pz).

For this reason, the comparison of pair (iii) was extended to all midline EEG

channels (Table 3.3b). Also in the additional channels (Fz and Cz), P300 could not

provide significant distinction for pairs of type (iii), whereas AvgLsum provided a

significant distinction with the largest effect in Cz (∆AvgLsum = 0.31 ± 0.09), in

accordance with the topographical map (see Figure 3.5, mid column, and Table 3.b).

3.3. Comparison of trialwise difference values (undistorted EEG recordings)

Test of difference against zero (intercept)

P300 AvgLsum

F p F p

Frequentbefore/rare Pz (31.60) Pz (<0.0005) Pz (101.18) Pz (<0.0005)

Rare/frequentafter Pz (25.63) Pz (0.001) Pz (80.36) Pz (<0.0005)

Table 3.3a. ANOVA repeated measures of the frequentbefore/rare and rare/frequentafter transitions

in electrode Pz. P-values of the intercept term are given, denoting significant deviation from

zero (df1=1 and df2=8).


Intercept

P300 AvgLsum Mean

amplitude in

W1

Mean absolute

amplitude in

W1

Frequentbefore/frequentafter F p F p F p F p

Fz

(0.01)

Cz

(0.000)

Pz

(1.16)

Fz

(0.718)

Cz

(0.999)

Pz

(0.313)

Fz

(3.60)

Cz

(11.33)

Pz

(6.71)

Fz

(0.094)

Cz

(0.010)

Pz

(0.032)

Fz

(10.90)

Cz

(13.28)

Pz

(2.87)

Fz

(0.011)

Cz

(0.007)

Pz

(0.130)

Fz

(0.02)

Cz

(1.17)

Pz

(2.90)

Fz

(0.891)

Cz

(0.311)

Pz

(0.127)

Table 3.3b. Undistorted EEG recordings. ANOVA repeated measures of the

frequentbefore/frequentafter transitions using the P300, AvgLsum, mean and mean absolute

amplitude of W1 (df1=1 and df2=8).

The grand average of the ERPs pointed out that amplitude differences

between frequentbefore and frequentafter were largest in a temporal window between

100 ms and 200 ms, which is covered by the time window W1 (used for AvgLsum),

however, not by the time window used for P300 detection. Thus, to avoid an overly

influence of different temporal windows analysed, the amplitude based analyses

(mean and mean absolute) of pairs of type (iii) for W1 were repeated (Table 3.3b).

This analysis revealed that the mean amplitude but not mean absolute amplitude of

W1 distinguished between subtypes of frequent tones, with the highest F-value in Cz

(∆mean amplitude of W1= -0.36 ± 0.10). Formally, the difference across AA and RQA

yielded a significant result (F(1,8) = 27.63, p=0.001). However, in the case of a

positive deflection in the electrophysiological data (as observed in W1), the difference

of amplitudes was negative, whereas the differences of AvgLsum were positive, which

may result in an artificial finding on the superiority of RQA. When considering this

difference of AvgLsum with respect to the negative numerical sign of the deflection (by


simply inverting the AvgLsum ∆ values), methods were not statistically different (F(1,8)

= 0.103, p = 0.757).

Cross-methods comparison of differentiation between rare and frequent tones

(EEG recordings inside the scanner)

MR image acquisition will introduce systematic noise in the simultaneous

measured EEG recordings. For the analysis of EEG data recorded inside the

scanner, the same channel (Pz) and measures (P300 versus AvgLsum of W1

[Equation (3.3)]) were employed as for the analysis of EEG data obtained outside

the scanner. Again, both methods discriminated between rare and frequent tones

with medium to large effect sizes (P300: 0.68 ± 0.20, t(8) = 3.42, p < 0.0005; AvgLsum

0.62 ± 0.11, t(8) = 5.65, p < 0.0005) (Table 3.4). AvgLsum showed higher effect sizes

in three of nine subjects, but the methods were not significantly different in a paired

comparison of their effect sizes (t(8)= 0.37 p = 0.723).

3.4. Effect sizes for AvgL and P300 derived from EEG recordings during an fMRI scan

Subject AvgLsum P300

1 0.52 -0.16

2 0.28 0.32

3 0.50 0.49

4 0.90 0.09

5 1.02 1.73

6 0.26 0.62

7 0.66 0.96

8 0.30 0.73

9 1.12 1.34


Table 3.4. Cohen’s effect sizes (rare vs. frequent) for P300 (channel Pz) and AvgLsum (channel

Pz, W1) for all subjects.

Cross-methods comparison of differentiation between subtypes of frequent

tones (EEG recordings inside the scanner)

Under the unfavourable EEG recording conditions during an MRI acquisition,

repeated measures ANOVA again revealed significant results for pairs of type (i) and

(ii) for both P300 and AvgLsum (Table 3.5a). For neither pair type, however,

differences between methods were significant (F(1,8) = 0.00, p = 0.987 and F(1,8) =

0.28, p = 0.612, respectively). For the pair type (iii), in accordance with the analysis

of the data recorded without concurrent MRI acquisition, AvgLsum and mean

amplitude of W1 and Cz were tested (Table 3.5b). Both measures distinguished

between frequentbefore and frequentafter tones (∆mean amplitude of W1 = -0.12 ± 0.05 and

∆AvgLsum = 0.22 ± 0.07). Yet, neither method was superior when compared under

consideration of the aforementioned insensitivity of RQA towards numerical

amplitude signature (F(1,8) = 1.01, p = 0.344).

3.5. Comparison of trialwise difference values (EEG recordings during an fMRI scan)

Intercept

P300 AvgLsum

F p F p

Frequentbefore/rare 11.88 0.009 33.22 <0.0005

Rare/frequentafter 13.31 0.007 47.00 <0.0005

Table 3.5a. ANOVA Repeated measures of the frequentbefore/rare, rare/frequentafter transitions in

electrode Pz. P-value of the intercept term is given, denoting significant deviation from zero

(d1=1, d2=8).


Intercept

AvgLsum Mean amplitude

of W1

Frequentbefore/frequentafter F p F p

Cz (9.16)

Cz (0.016)

Cz (6.80)

Cz (0.031)

Table 3.5b. ANOVA repeated measures of the frequentbefore/frequentafter, transitions using the

AvgLsum, mean amplitude of W1 (df1=1 and df2=8).

Correlation between EEG amplitude based measures and AvgLsum

Linear correlations between P300 and AvgLsum (Figure 3.7a) were weakly

positive (maximum r2 = 0.30) and significant for all subjects (p < 0.01). The

correlation between the mean amplitude of the ERP signal in W1 and AvgLsum was

extremely low (r2 < 0.04), with only one individual subject showing a significant

correlation (r2 = 0.032, p<0.01). However, as mentioned above, AvgLsum can just be a

positive number while the mean amplitude may also show negative values. When

mean absolute amplitudes are used to avoid spanning of the variance by negative

amplitude values, correlations were more strongly positive and again significant for all

subjects (p < 0.01) (Figure 3.7b).


Figure 3.7. Single trial correlation between AvgLsum and AA measures. (a) Weak positive

correlations (r2 between 0.044 and 0.308) between P300 amplitude (µV) and AvgLsum were

found. (b) Similarly, the use of the same time window (W1) for RQA and AA resulted in weak

correlations between AvgLsum and the mean absolute amplitude (r2 between 0.040 and 0.449).

Chapter 3 99

Discussion

In this study, RQA was applied to analyze EEG recorded during a classical

auditory oddball experiment. Its power to distinguish between different tone types

was compared to amplitude based analysis. Furthermore, the robustness of RQA

towards unfavourable EEG recording conditions as typically found during

simultaneous recordings of EEG and MRI was investigated. In brief, the main findings

were: (1) RQA could robustly detect differences in the non-linear signal structure

between rare and frequent tones. (2) This discrimination ability was maintained for

the EEG recorded with concurrent MRI acquisition. (3) Compared with amplitude

based analysis, no statistically significant difference was found regarding the ability to

distinguish between tone types. 4) Linear correlations between the two methods

across trials were low to medium, supporting that RQA may extract ERP features that

are different from simple amplitude characteristics.

Previous attempts of ERP single trial characterization

A number of alternative approaches have been used in literature to address

the problem of single trial ERP characterization before: Bénar et al. [152] tracked

trial-by-trial amplitude changes of the EEG response recorded during an oddball

paradigm. Both at the single subject level and for an fMRI group analysis, the authors

found that ERP amplitudes explained a significant amount of variability of the fMRI

BOLD response. Mantini et al. [157] used the EEG to measure P300 amplitude

variability across trials. Other than Bénar et al. [152], they compared the individual

tone trials with an average-based P300 template of the rare tones and forwarded this

comparative value to fMRI analysis. They reported that such modelling of the trial-by-

trial variability correlates with dorsal and ventral attention networks, providing


conceptual validation that trial-by-trial variability indeed carries important information

on the current cognitive state of a subject, e.g. with respect to attention. Demiralp et.

al. [177] described a wavelet based method to characterize single trials of an oddball

experiment. In their analysis, a selected wavelet coefficient correlated with the EEG

delta frequency power was defined as a classifier to group ERPs into P300 or non-

P300 trials. They found that the formation of a P300 response was more robustly

predicted by the wavelet based classifier than the rare/frequent classification. By

applying ICA to EEG single trials simultaneously recorded with fMRI, Debener et al.

[153], reported that the EEG correlate of performance monitoring predicted the BOLD

response in the anterior cingulate and adjacent medial prefrontal cortex. Bagshaw

and Warbrick [155] compared wavelet based and ICA based denoising techniques,

demonstrating that wavelet denoising (WD) was superior to ICA. Zouridakis et al.

[156] proposed an iterative ICA (iICA) to enhance specific components of single trials

ERPs. Iyer and Zouridakis [154] showed that iICA improves the detection of ERP

components in single trials if compared with WD and classical averaging. However,

one of the general problems of ICA is that components of true neuronal origin may

falsely be dismissed, leading to an underestimation of the true inter-trial neuronal

variability. One common criterion to select components is a minimum correlation

between the component and the ERP average. As pointed out by Bagshaw and

Warbrick [155], this procedure can also reduce the contribution of non- phase-locked

activity. In this respect, RQA is more explorative as it is not based on any assumption

on the structure of the response signal, including the assumptions that ERP activity is

phase-locked. Previous reports applied RQA to detect minute changes in EEG event-

related responses at the level of individual trials [164;165;178;179]. These studies,


however, lack formal statistical cross-methodological comparisons, for example with

amplitude based ERP analyses.

Comparison of RQA and AA based on pooled trial responses

When trial responses were pooled during a complete auditory oddball

experiment, both RQA (represented by AvgL) and AA (represented by the P300

amplitude) provided significant distinction between the two main types of tone trials.

For both methods, this distinction proved significant even at the single subject level,

and also when EEG recordings were distorted by fMRI related artifacts. When

individual effect sizes of RQA and AA were compared against each other in a

pairwise manner, no significant difference emerged. As would be expected, MR-

scanner related distortions led to slight general decrease of effect sizes, however,

again both methods performed equally well in differentiating tones. Given the overall

power of the present analysis, the results do not support a generally higher sensitivity

of RQA towards electrophysiological signal changes as induced by tone deviance in

the auditory oddball paradigm.

While there is a huge amount of literature pinpointing the functional

significance of specific ERP deflections, e.g. the P300 [145;147;148;166;167;169],

the RQA method, still being in its infancy as applied on biosignals like EEG, does not

provide an easily understandable neurophysiolocal correlate. In ERP studies, P300 is

known to be particularly evoked by rare stimuli, with its amplitude increasing from

midfrontal to midparietal brain areas [180;181]. When RQA was applied for signal

analysis in a time window that included the P300 component, a spatial maximum of

the RQA variable AvgLsum over parietal midline electrodes was observed, similar to

the typical maximum observed in ERP studies.


AvgL, DET and MaxL are RQA variables all based on diagonal lines features

of the RP. Not surprisingly, these variables showed high linear correlation with each

other. Generally, diagonal lines in the RP represent similar sequential pattern of

states at different time points, reflecting information on the system’s divergence.

AvgL in particular reflects the average duration of such a sequential pattern, and can

thus be interpreted as the mean prediction time of the system [160]. In all subjects,

AvgLsum was higher for rare tones than for frequent tones, suggesting a less

divergent system in response to stimulus deviance compared with the frequent

stimulus.

Comparison of RQA and AA based on comparisons of subsequent trials

Pooling all trials of a whole oddball experiment may lead to underestimation

of differences between the analysis methods. Therefore, the analysis as presented

here was expanded in two ways: first, focusing on subsequent pairs of trials, which

may induce fluctuations in the cerebral response, for example, due to altered

vigilance state or habituation; second, pairs of corresponding differential responses

were compared across the analysis methods. While again demonstrating that both

methods are capable to detect strong response changes induced by tone deviance,

this comparison confirmed results of the analysis of pooled trials with no significant

superiority of either method in terms of sensing tone deviance.

With respect to discrimination of sub-types of frequent tones (frequentbefore vs

frequentafter), it was expected that the neuronal response may depend on the specific

brain state prior to stimulus delivery. Analyzing the influence of the frequent tone

position relative to the rare tone, Hirata and Lehmann [182] have previously reported

that the absolute amplitude of the N100 deflection was stronger for frequent tones


preceding rare tones than for frequent tones following rare tones. Starr et al. [183]

showed that this decrease of N100 after rare tone presentation was largest for the Cz

electrode position. They suggested that this N100 decrease corresponds to memory

updating processes required after disruption of a series of regular tones by a deviant

tone.

The a priori hypothesis of the present work was that RQA could be more

sensitive in detecting such brain state differences, potentially encoded in differences

of signal regularity rather than amplitude averages. Based on the undistorted EEG

recordings, mean amplitude of W1 but not P300 could distinguish between subtypes

of frequent tones. In fact, amplitude values within the W1 time window were higher

for frequentafter compared with frequentbefore values. This indicates that the influence

of the N100 response contained in W1 could have led to this result. To confirm this,

the N100 of frequentafter and frequentbefore trials was explicitly analyzed, showing

significant difference into the expected direction (F(1,8) = 5.45, p = 0.048).

Comparing N100 and mean amplitude of W1, no difference was seen in the trialwise

comparison frequentafter/frequentbefore differences of both measures (F(1,8) = 0.52, p =

0.490). This result confirmed that previously reported N100 differences were mainly

underlying the amplitude difference between frequentafter and frequentbefore found for

W1. In the other hand, RQA resulted in higher AvgLsum values for frequentbefore than

frequentafter. When this difference in the sign of de signal was considered

appropriately, RQA again was not significantly better than AA in differentiating

frequentbefore and frequentafter. Similar results of the comparison between subtypes of

frequent tones were found for data recorded with concurrent MRI acquisition. The

result pattern as a whole suggests that differentiation between the two states seems

to depend rather on inclusion of a critical early time window around 100 - 200 ms


than on the different analysis methods. To define whether the extraction of features

from a single trial provides a more valid surrogate of the response to the rare event,

parallel independent measures of brain activation such as BOLD-fMRI are needed.

Influence of embedding parameters and other parameters on RQA

performance

Following Taken’s theorem [171], an embedding dimension higher than twice

the dimension of the data set should be used in RQA. However, Iwanski and Bradley

[184] showed that for certain low-dimensional systems, it is possible to obtain similar

results without embedding the raw data. Thus, the choice of ‘optimal’ embedding

parameters for a specific application is rather arbitrary. In the present study, a sum

score of several m and d values was used. In doing so, the analysis was not

restricted to a given set of m and d, but considered a broader range of possible

embedding parameters, allowing for a more generalized evaluation of the RQA

approach. Also, inter-subject variability which may be inherent to physiological data is

minimized by such an approach. However, the noise may affect RQA if non-optimal

embedding dimensions are chosen, leading to an underestimation of RQA

performance [175]. In the EEG recordings, the currents induced by rapidly switching

magnetic field gradients during the concurrent MRI acquisition introduce noise and

likely change the complexity of the data structure. Therefore, using a range of

multiple values of m and d may overemphasize such noise components. Hence, an

analysis using only a single pair of m and d parameters optimized for the individual

experiment is presented in the Appendix A3, Table A3.1 shows the embedding

parameters associated with the highest effect size per subject for optimal EEG

recordings and for EEG recordings distorted by concurrent MRI. Even when


optimizing the RQA parameter for each subject, the results did not improve rare and

frequent tones discrimination if compared with P300 amplitude analysis (paired t-test

of individual effect sizes, t(8) = -0.77, p = 0.462 and t(8) =1.80, p = 0.311 for data

with and without concurrent MRI acquisition, respectively). In a previous work, RQA

was compared with ERP analysis, however in the context of a different stimulus

paradigm. Schinkel et al. [164], when analysing the ERP response to semantic

mismatch, showed that RQA could classify about 14% of the trials correctly while

EEG succeeded in only 1.5%. However, no formal statistical comparison between the

two methods has been provided, and EEG preprocessing seemed not to be

optimized for ERP amplitude analysis.

In the present work, the focus of the RQA was set on optimizing the

embedding parameters m and d , however, RQA performance depends on many

additional factors such as minl , minv , threshold and norm used to define the distance

between two states (see Table 3.1). Marwan et al [179] suggested that RQA based

on order patterns, in comparison with FAN, would better reveal the P300 component.

Preliminary analyses comparing the power of the RQA using FAN and using order

pattern (not shown) revealed that results were better when RQA was performed

using FAN. This leaded the present analysis to be based on FAN.

Correlation between EEG amplitude based measures and AvgLsum

P300 amplitudes and AvgLsum were significantly positively correlated in all

subjects (Figure 3.7a). While significance was mostly based on the large number of

trials, the explained variance (r2) was relatively low. This may point out that RQA

extracts a type of information different from P300 amplitude based information.

However, one needs to consider that RQA covers a much larger temporal window


(W1) than the window used for P300 extraction. Still, the correlation between mean

amplitude of W1 and AvgLsum was extremely low (r2<0.030). As a measure of

temporal recurrence, AvgLsum is insensitive to the numerical sign of the signal, i.e. it

does not differentiate positive and negative EEG deflections of the same amplitude

and shape. Therefore, it is clear that the mean absolute amplitude of W1 and AvgLsum

showed much higher correlation. This feature of RQA parameters is also reflected in

the comparison between frequentbefore and frequentafter tone types: Here, AvgLsum

showed higher values for frequentbefore while N100 and mean amplitude of W1

showed higher values for frequentafter. In summary, according to the theoretical

background of RQA, it may provide additional information on the electrophysiological

response to an auditory oddball paradigm, however, further studies are required to

determine the functional correlate of AvgLsum and also to ensure that AvgLsum is not

just reflecting linear changes in signal amplitude. For example, this could be tested

by exploiting experiments with variable tone probabilities, different inter-stimulus

intervals and different tone frequencies and amplitudes, all known to modulate the

electrophysiological response in ERP experiments.

Conclusions

In this study, a formal statistical comparison of the performance of RQA and

classical amplitude based ERP analysis was provided. RQA provided a robust

distinction between the two main types of trials in an auditory oddball experiment

(frequent and rare tones), as well as between different subtypes of frequent tones.

Minor signal corruption was observed in the presence of MRI artifacts, however, both

analysis methods were affected to the same degree. Despite similar discrimination

power, RQA may extract a qualitatively different type of information from the EEG


signal, as suggested by the theoretical concept of RQA. Hence, it still remained

unclear if RQA should be preferred over ERP analysis, especially when considering

the computational effort and the conceptual complexity of the method. Still, RQA may

be used as a complementary tool in a single-trial analysis. While the physiological

significance of ERP amplitudes has been studied in depth, a direct neuronal correlate

of RQA parameters has not yet been defined. Particularly the comparison of

combined fMRI/EEG data with variable tone probabilities, different inter-stimulus

intervals and different tone frequencies and amplitudes may be useful to elucidate

the different neural correlates of amplitude versus complexity measures such as

provided by RQA.

108

CONCLUSION AND OUTLOOK

In this thesis, the potential of simultaneous EEG/fMRI is exploited with

different methodological approaches. The basic principles and current state-of-the-art

analysis methods of EEG and fMRI are outlined (Chapter 1). Such complementary

data were used to address two major research questions: To investigate alterations

in cerebral functional connectivity between the hippocampus and neocortex during

wakefulness and different sleep stages (Chapter 2), and to evaluate the potential of

recurrence quantification analysis to electrophysiological data (Chapter 3).

In Chapter 2, fMRI time-series of subregions of the hippocampal formation

(HF) (cornu ammonis, dentate gyrus and subiculum) were extracted based on

cytoarchitectonical probability maps. We observed sleep stage dependent changes in

HF functional coupling. The HF was integrated to a variable extent in the default

mode network (DMN) in wakefulness and light sleep stages, but not in slow wave

sleep. The strongest functional connectivity between the HF and neocortex was

observed in sleep stage 2 (compared with both slow wave sleep and wakefulness).

We observed a strong interaction of sleep spindle occurrence and HF functional

connectivity in sleep stage 2, with strongest HF/neocortical connectivity during

spindles. Moreover, the cornu ammonis reveals strongest functional connectivity with

the DMN during wakefulness, while the subiculum dominates connectivity to frontal

brain regions during sleep stage 2. Increased connectivity between HF and

neocortical regions in sleep stage 2 suggests an increased capacity for global

information transfer, while connectivity in slow-wave sleep is reflecting a functional

system optimal for segregated information (re-)processing. The present data may be

relevant to differentiating sleep stage specific contributions to neural plasticity as

proposed in models of sleep dependent memory consolidation.

Conclusion and outlook 109

In Chapter 3, the potential of recurrence quantification analysis (RQA) is

evaluated with regards to improvements in the analysis of trial-by-trial-variability in

event-related potential (ERP) experiments. For this propose, EEG data were

acquired during an auditory oddball paradigm to compare the efficiency of RQA with

amplitude based analysis of single trial ERPs. These methods were compared with

regard to the power to distinguish between frequent and rare tones and to distinguish

between subsequent trials. Further, the robustness of both methods was evaluated

towards structured noise induced by simultaneous magnetic resonance imaging

(MRI). RQA provided robust discrimination of rare and frequent tones and of

subsequent trials based on their different signal structure. This discrimination ability

was maintained when the EEG recordings were obtained in the MR scanner. In terms

of power to discriminate between rare and frequent tones, RQA was not significantly

superior to conventional amplitude analysis. Despite equal discrimination power,

RQA measures were only weakly correlated with ERP amplitudes, suggesting that

additional information on single trials may be extracted by RQA. Nevertheless, it still

remains unclear if RQA is to be preferred over ERP analysis, especially when

considering the computational effort and conceptual complexity of the method. Still,

RQA has the potential of detecting small differences in response to identical tones

when presented in a changed/altered context, i.e. dependent on their relative position

to the last deviant tone. According to the present results, RQA may be used as an

additional tool to obtain new insights in neurophysiological and neuroanatomical

correlates of stimulus habituation and novelty processing even in combined trial-by-

trial functional MRI/EEG experiments.

Nowadays, several different imaging approaches have been proposed as a

tool to study metabolic, hemodynamic or electromagnetic functional activity. The


main focus of these methodologies is to identify the localization of the neuronal

activity. However, the information of how the brain regions communicate with each

other is also crucial to understand the functional organization of cortical regions.

Therefore, we applied a seed analysis to quantify connectivity of the hippocampus

with the neocortex during different sleep stages. In 2005, Eguiluz et al. [185]

proposed a method to extract functional networks using fMRI and analyze them in the

context of the current understanding of complex networks. In this method, the

connectivity between brain regions is measured using a binary correlation matrix. In

this way, a functional connectivity is established if the temporal correlation between

two specific regions exceeds a positive threshold. Based on this matrix, statistical

properties of the network as e.g. the efficiency and the ‘cliqueness’ can be

calculated. Recently, this method was applied by our group to explore how

physiological changes during sleep were reflected in functional connectivity and

network properties [114]. This study leaded to insights about changes in

consciousness in the descent to sleep and the capacity of the brain to integrate

information across functional modules during sleep. Linear correlations are the most

used measure of connectivity, however, it can miss an important part of functional

dependence as coupling among brain regions should be considered nonlinear [186].

Therefore, one of the goals of my PhD studies was to evaluate the potential of the

nonlinear method RQA to be applied to electrophysiological data and test the

robustness of this method towards structured noise induced by simultaneous

magnetic resonance imaging, since an important aim in neuroscience is to

successfully combine fMRI and EEG recordings. In our analysis we applied RQA for

three individual EEG channels, analysing changes in recurrence patterns and

deterministic behaviour in data recorded during an oddball experiment. Indeed, we


also tried to apply RQA in sleep data, as well as in a modified oddball protocol with

subjects just passively listening to tones. Although RQA lacked to show superiority in

detecting changes in the EEG as compared with a linear amplitude based method,

RQA likely provides different information, considering the weak correlation with linear

based values. The recurrence plot may still be used to study generalized

synchronization between channels to capture additional nonlinear changes in the

neuronal networks. In 1995, Rulkov et al. [187] defined that a generalized

synchronization exist "when trajectories in the phase spaces of driving and response

systems are connected by y(t) =Ψ (x(t)), two close states in the phase space of the

response system correspond to two close states in the space of the driving system".

This could be simply calculated by the products of the recurrence plot of two different

channels [188]. Linear and nonlinear changes of brain networks regarding different

levels of arousal could therefore be investigated. A preliminary analysis suggested

that this approach may be more powerful in detecting electrophysiological changes

than the use of a single channel and also than synchronization by means of linear

correlations. RQA could, therefore, still hold promise to examine the brain

connectivity.

112

APPENDIX A2

A2.1. HF functional connectivity during wakefulness

Brain Region

Brodmann

areas,

deep nuclei

Cluster size

(voxel)

Peak voxel

T-value MNI

coordinates

1 L/C/R

(Posterior) cingulate, uncus, fusiform gyrus, precuneus, culmen, claustrum, thalamus, hippocampus, amygdala

R7,L20,23, 27-31,34-36,L37

7548 17.37 28 -28 -

16

2 R Inferior/middle temporal gyrus

21 322 10.93 50 2 -

32

3 R Middle/superior frontal gyrus 8 207 9.72 24 22 54

4 L/R Anterior cingulate, medial frontal gyrus

10,32,L42 886 9.40 6 56 14

5 L Angular gyrus 39 305 9.37 -44 -70 30

6 L Middle/superior frontal gyrus 8 187 9.23 -22 28 52

7 R Middle/superior temporal gyrus, angular gyrus, supramarginal gyrus

39 491 8.81 44 -64 28

8 L Middle/superior temporal gyrus

21 298 8.25 -56 -14 -

14

Table A2.1. Clusters resulting from second level random effects analysis (pFWE <10-6, extent >

150). Regions showing significant functional connectivity with all HF subregions are listed.

Sorting is after T-values of the cluster peak voxel. Brodmann areas are identified for clusters

covering > 3% of the respective area. Coordinates are given in MNI space. L/R/C denote the

left/right hemisphere or central position, respectively.

Appendix A2 113

A2.2. Sleep stage specific functional HF connectivity changes

Brain Region Brodmann areas,

deep nuclei

Cluster size

(voxel)

Peak voxel

T-value MNI

coordinates

Wakefulness > S1

1 L/R Thalamus Ventral anterior nucleus, medial dorsal nucleus

270 4.36 4 -18 8

Wakefulness < S1

1 R Fusiform gyrus, culmen 291 5.08 44 -38 -24

2 L/C/R Middle/superior occipital gyrus, cuneus

R17-R19 858 4.69 16 -82 30

3 L Culmen 152 4.47 -22 -44 -18

4 R Middle temporal gyrus 37,39 271 4.46 42 -68 8

5 L/C/R Precuneus, paracentral lobule

5,L7 463 4.45 -4 54 64

6 L Middle/superior temporal gyrus

41 99 4.11 -52 -28 2

Wakefulness > S2

No suprathreshold cluster

Wakefulness < S2

1 L

Transverse/middle/superior temporal gyrus, inferior frontal gyrus, insula, inferior parietal lobule

13,21,22,29,38, 40-42,44,45,47

2367 6.85 -64 -8 -4

2 L/C/R

Posterior cingulate, fusiform gyrus, inferior occipital gyrus, lingual gyrus, cuneus

17,18,L23,30 2564 6.81 8 -88 -8

3 R Transverse/middle/superior temporal gyrus, postcentral gyrus

21,22,41-43,47 1221 5.40 68 -18 8

4 L Precentral gyrus 1,3,4,6 381 5.12 -46 -6 56

5 L/C/R Cingulate gyrus, paracentral lobule, medial frontal gyrus

23,R24,L31 293 4.68 6 -12 36

Appendix A2 114

Wakefulness > SWS

1 L Angular gyrus, supramarginal gyrus, inferior parietal lobule

39,40 659 5.55 -48 -70 30

2 R Middle frontal gyrus 89 5.50 42 52 -6

3 R

Angular gyrus, supramarginal gyrus, inferior/superior parietal lobule

39,40 628 4.83 46 -66 28


10,L32 744 4.77 10 48 -10

5 L Inferior/medial frontal gyrus

25,47 119 4.64 -20 20 -20

6 L/C/R (Posterior) cingulate, precuneus

L7,L23,L30,L31 300 4.10 -4 -52 24

7 L Inferior/middle frontal gyrus

47 60 3.91 -32 32 -18

8 R Inferior/middle frontal gyrus

43 3.84 36 34 -12

9 R Inferior/middle/medial fronatl gyrus

39 3.77 22 24 -18

Wakefulness < SWS


S2 > SWS

1 R Transverse/middle/superior temporal gyrus, inferior frontal gyrus

21,22,29,41,47 727 5.30 66 -12 -6

2 L Inferior/medial frontal gyrus

25,47 249 4.89 -28 24 -20

3 L Transverse/superior temporal gyrus, inferior frontal gyrus

21,22,41-43,45,47 827 4.71 -58 -22 8

4 L/C/R Posterior cingulate, lingual gyrus, cuneus

30 647 4.40 8 -88 -8

Appendix A2 115

5 L/R Orbital gyrus 11 263 4.39 8 52 -20

6 L/C Precuneus 5,7 454 4.33 -4 -52 64

7 L/R Superior/medial frontal gyrus

L9 197 4.27 -2 58 32

8 R Pre-/postcentral gyrus 3,4 462 4.20 62 -10 30

S2 < SWS


Table A2.2. Clusters resulting from second level random effects analysis (pFWE,cluster<0.05,

collection threshold puncorr<0.001). Regions showing significant functional connectivity with all

HF subregions are listed. Sorting is after T-values of the cluster peak voxel. Brodmann areas

are identified for clusters covering > 3% of the respective area. Coordinates are given in MNI

space. L/R/C denote the left/right hemisphere or central position, respectively.

Appendix A2 116

A2.3. Subregional HF functional connectivity per sleep stage.


deep nuclei

Cluster size

(voxel)

Peak voxel

T-value MNI

coordinates

Wakefulness

CA

1 R

Sub-gyral, uncus, fusiform gyrus, inferior temporal gyrus, extra-nuclear, lentiform nucleus, claustrum, caudate, hippocampus, amygdala, lateral globus pallidus, putamen

19-21,27,28,34-39 4151 12.30 32 -10 -20

2 L

Uncus, fusiform gyrus, middle occipital gyrus, culmen, extra-nuclear, lentiform nucleus, claustrum, caudate, hippocampus, amygdala, lateral globus pallidus, putamen

20,28,34-37 3063 11.91 -26 -10 -20

3 L Inferior/middle temporal gyrus

21, 22 702 5.48 -58 -24 -10

4 L Parahippocampal gyrus, thalamus

27 99 5.22 -14 -38 4

5 R putamen 67 4.88 22 4 6

6 L Inferior frontal gyrus 47 350 4.85 -30 34 -12

7 L/R Thalamus Medial/lateral dorsal nucleus, midline nucleus

196 4.60 -12 -12 18

8 L/C/R Anterior cingulate 25 101 4.46 2 8 -8

9 R Insula, claustrum 13 132 4.36 40 -10 16

10 L/C Posterior cingulate 23,29,30,31 230 4.35 -4 -52 18

11 R Superior parietal lobule 71 4.31 24 -48 58

12 C Midbrain 47 4.31 -2 -26 -18


32 330 4.29 0 58 6

14 R Middle/superior temporal gyrus

56 4.25 70 -38 8

Appendix A2 117

15 L Superior frontal gyrus 74 4.12 -14 50 40

16 R Precentral gyrus 64 4.12 56 0 30


39 69 4.10 52 -60 24

18 L Angular gyrus 39 144 3.91 -40 -66 26

19 R Inferior/middle frontal gyrus

53 3.78 38 36 -14

20 R Middle temporal gyrus 21 67 3.76 66 -22 -10

S1

CA

1 L

Uncus, fusiform gyrus, culmen, hippocampus, amygdala, caudate tail, lateral globus pallidus, putamen

20,28,34-37 2463 12.31 -32 -14 -20

2 R

Sub-gyral, uncus, fusiform gyrus, inferior/middle temporal gyrus, culmen, claustrum, hippocampus, amygdala, caudate tail, lateral globus pallidus

20-22,28,34-38 4055 11.36 36 -16 -18

3 L Inferior/middle/superior temporal gyrus

21,38 1120 4.89 -54 -26 0

4 R Inferior/middle temporal gyrus

37,39 294 4.82 58 -66 -2

5 L/C Precuneus, paracentral lobule

5 375 4.81 -4 -44 62


21,22 195 4.66 68 -36 0

S2

CA

1 L/C/R

Sub-gyral, uncus, fusiform gyrus, inferior occipital gyrus, inferior temporal gyrus, insula, lingual gyrus, middle/superior temporal gyrus, extra-nuclear, lentiform nucleus, claustrum, hippocampus,

13,19,20-22,28, 34-38,42

6066 10.62 26 -12 -16

Appendix A2 118

amygdala, caudate tail, putamen, lateral globus pallidus

2 L

Sub-gyral, transverse temporal gyrus, uncus, fusiform gyrus, inferior occipital gyrus, inferior temporal gyrus, insula, lingual gyrus, middle occipital gyrus, middle temporal gyrus, culmen, claustrum, hippocampus, amygdala, caudate tail, lateral globus pallidus, putamen

19-22,28,34-38,41, 42

5796 10.56 -34 -12 -22

3 L Lingual gyrus 17 193 5.38 -6 -92 -8

4 L Superior temporal gyrus, supramarginal gyrus, inferior parietal lobule

13,40 415 4.99 -54 -46 20

5 L Inferior frontal gyrus 47 280 4.9 -40 28 -14

6 R Lingual gyrus 145 4.31 12 -82 -16

7 R Claustrum, lentiform nucleus

174 4.26 -24 16 2

8 L Precentral gyrus 201 4.18 -60 0 32

DG

1 R Inferior temporal gyrus 76 4.72 52 -24 24

2 R Thalamus, hippocampus 27 165 7.57 22 -36 4

3 L Thalamus, hippocampus 27 130 7.10 -24 -36 0

SUB

1 L/R Anterior cingulate, superior/medial frontal gyrus

9,10,32 1892 4.69 8 40 8

2 L Middle/superior frontal gyrus

247 3.91 -30 40 20

SW

Appendix A2 119

CA

1 R

Sub-gyral, uncus, fusiform gyrus, middle/superior/inferior temporal gyrus, hippocampus, amygdala, caudate tail

20,21,28,34-36,38 2834 11.15 34 -18 -16

2 L

Sub-gyral, uncus, fusiform gyrus, inferior temporal gyrus, middle/superior temporal gyrus, hippocampus, amygdala, caudate, lateral globus pallidus

20- 22,28,34-37 3162 10.99 -26 -10 -22

3 L Superior/inferior frontal lobe

38 138 4.80 -44 24 -18

DG

1 R Thalamus 27 111 5.51 26 -32 0

2 L/C/R Posterior cingulate, cuneus, precuneus

18,31 486 5.46 -20 -68 14

SUB

1 R Parahippocampal gyrus 28,35 241 4.78 28 -8 -32

2 R Superior/medial frontal gyrus

127 4.31 14 38 46

3 L Pre-/postcentral gyrus 2-4 279 4.24 -34 -28 54

4 R Middle/superior frontal gyrus

6 142 3.97 26 8 60

Table A2.3. Clusters resulting from second level random effects analysis (pFWE,cluster<0.05,

collection threshold puncorr<0.001). Regions showing significant stronger functional

connectivity with the specified HF subregions as compared to the other two subregions are

listed (cornu ammonis, CA; dentate gyrus, DG; subiculum, SUB). Sorting is after T-values of

the cluster peak voxel. Brodmann areas are identified for clusters covering > 3% of the

respective area. Coordinates are given in MNI space. L/R/C denote the left/right hemisphere or

central position, respectively.

Appendix A2 120

A2.4 Activity related to fast sleep spindles (14 - 16 Hz)


deep nuclei

Cluster size

(voxel)

Peak voxel

T-value MNI

coordinates

1 L/C/R

Cingulate, posterior/anterior cingulate, precuneus, paracentral lobule, postcentral gyrus, superior/medial frontal gyrus

5-7,10,23,24,29, 31,32,42

5588 4.89 10 26 28

2 R Transverse/superior temporal gyrus, insula

13,22,41,42 569 4.77 58 -20 10

3 L/C/R Lingual gyrus, cuneus 17,18 925 4.71 14 -98 0

4 L Pre-/post-central gyrus 2-4 440 4.66 -30 -28 64

5 R Superior frontal gyrus 9 345 4.50 26 44 46

6 L Transverse temporal gyrus* 41 103 4.12 -44 -30 10

7 R Thalamus 195 4.10 6 -8 4

8 R Postcentral gyrus 3-5 225 3.86 32 -32 62

9 L Supramarginal gyrus* 116 3.79 -54 -46 30

Table A2.4. Clusters resulting from second level random effects analysis (pFWE,cluster < 0.05,

collection threshold puncorr <0.001). Sorting is after T-values of the cluster peak voxel.

Brodmann areas are identified for clusters covering > 3% of the respective area. Coordinates

are given in MNI space. L/C/R denotes left/central/right clusters. (*) marks two clusters which

do not reach statistical significance, but only trend values pFWE,cluster < 0.1. L/R/C denote the

left/right hemisphere or central position, respectively.

Appendix A2 121

A2.5. Activity related to the interaction effect of Spindle × SUB


deep nuclei

Cluster size

(voxel)

Peak voxel

T-value MNI

coordinates

1 L/C/R

(Posterior) cingulate, Subcallosal gyrus, sub-gyral, transverse/inferior/middle/ superior temporal gyrus, uncus, fusiform gyrus, insula, lingual gyrus, pre-/post-central gyrus, paracentral lobule, inferior/middle/medial frontal gyrus, inferior/superior parietal lobule, precuneus, culmen, thalamus, extra-nuclear, lentiform, claustrum, hippocampus, amygdala, putamen

1-7,9,13,19,20-24, 28-47

21386 14.15 32 -22 -22

2 L

Transverse/middle/superior temporal gyrus, insula, pre-/post-central gyrus, inferior frontal gyrus, inferior parietal lobule, extra-nuclear, claustrum, putamen

6,9,13,22,29,38, 41-45,47

5279 11.86 -56 2 -6

3 L Thalamus, hippocampus, amygdala

27,28,30,34-37 1288 9.56 -24 -20 -20

4 L/C Inferior/middle occipital gyrus, cuneus

18 404 7.83 -24 -98 6

5 R Inferior occipital gyrus 329 7.08 34 -76 -10

6 L Superior frontal gyrus 8,9 518 6.40 -26 38 48


10,32,42 897 6.24 2 48 12

8 R Superior frontal gyrus 8,9 387 5.81 14 46 52

Table A2.5. Clusters resulting from second level random effects analysis (puncorr<10-4,

extent>150 voxel). Sorting is after T-values of the cluster peak voxel. Brodmann areas are

identified for clusters covering > 3% of the respective area. Coordinates are given in MNI

space. L/C/R denotes left/central/right clusters.

Appendix A2 122

A2.6. Activity related to the interaction effect of Spindle × FD

Brain Region

Brodmann

areas,

deep nuclei

Cluster

size

(voxel)

Peak voxel

T-

value

MNI

coordinates

1 L/R

Posterior cingulate, sub-gyral, transverse/inferior/middle/superior temporal gyrus, fusiform gyrus, inferior/middle occipital gyrus, insula, lingual gyrus, pre-/post-central gyrus, inferior/superior/medial frontal gyrus, cuneus, angular gyrus, inferior parietal lobule, paracentral lobule, declive, culmen, extra-nuclear, claustrum, thalamus, hippocampus

1-7,9,13,17-19, 21-24,27-31,

35-44 22287 14.59 -22 -34 -4

2 L/C Inferior/middle occipital gyrus, Lingual gyrus, cuneus

17, 18, 19 1006 8.26 -22 -

102 4

3 L Fusiform gyrus 37 248 6.25 -44 -46 -

22

4 L Inferior temporal gyrus 37 188 5.42 -54 -74 4





Appendix A2 123

A2.7. Activity related to the interaction effect of Spindle × CA


deep nuclei

Cluster size

(voxel)

Peak voxel

T-value MNI

coordinates

1 L

Fusiform gyrus, lingual gyrus, middle occipital gyrus, cuneus, hippocampus, amygdala

17-20,28,34-37 1581 8.97 -26 -28 -16

2 R

Uncus, fusiform gyrus, inferior temporal gyrus, culmen, hippocampus, amygdala

19-22,27,28,34-37 2803 8.42 22 -12 -24

3 L Precentral gyrus 43 226 7.65 -58 0 28

4 L Superior temporal gyrus 22 310 6.78 -46 -48 16

5 L Superior temporal gyrus 22 418 6.31 -58 -2 -4

6 C Paracentral lobule, medial frontal gyrus

186 6.23 4 -14 50

7 R Middle temporal gyrus 232 5.97 36 -68 -2





124

APPENDIX A3

A3.1 Higher effect sizes found for each subject with optimized m and d values

Undistorted EEG recordings Distorted EEG recordings

during an fMRI scan

Subjects m d Effect size m d Effect size

1 3 10 1.03 5 8 0.68

2 3 14 1.28 11 3 0.63

3 4 11 0.97 5 12 0.70

4 3 10 1.48 4 9 0.88

5 3 16 1.26 2 10 1.19

6 4 7 0.98 7 6 0.78

7 3 16 1.03 3 7 0.76

8 3 16 1.00 5 8 0.52

9 3 18 0.77 5 7 1.08

Table A3.1a. Cohen’s effect sizes (rare vs. frequent tones) for m and d individually optimized

for each subject.

125

REFERENCES

[1] Dolan,R.J., Neuroimaging of Cognition: Past, Present, and Future.

Neuron 60 (2008) 496-502.

[2] Buxton,R.B., Interpreting oxygenation-based neuroimaging signals: the

importance and the challenge of understanding brain oxygen metabolism. Frontiers in

Neuroenergetics 2 (2010) e8.

[3] Logothetis,N.K., Wandell,B.A., Interpreting the BOLD signal. Annual

Review of Physiology 66 (2004) 735-769.

[4] Shibasaki,H., Human brain mapping: Hemodynamic response and

electrophysiology. Clinical Neurophysiology 119 (2008) 731-743.

[5] Ritter,P., Villringer,A., Simultaneous EEG-fMRI. Neuroscience &

Biobehavioral Reviews 30 (2006) 823-838.

[6] Bandettini,P.A., What's New in Neuroimaging Methods? Year in

Cognitive Neuroscience 2009 1156 (2009) 260-293.

[7] Czisch,M., Wetter,T.C., Kaufmann,C., Pollmächer,T., Holsboer,F.,

Auer,D.P., Altered processing of acoustic stimuli during sleep: Reduced auditory

activation and visual deactivation detected by a combined fMRI/EEG study.

Neuroimage 16 (2002) 251-258.

[8] Laufs,H., Kleinschmidt,A., Beyerle,A., Eger,E., Salek-Haddadi,A.,

Preibisch,C., Krakow,K., EEG-correlated fMRI of human alpha activity. Neuroimage

19 (2003) 1463-1476.

[9] Czisch,M., Wehrle,R., Kaufmann,C., Wetter,T.C., Holsboer,F.,

Pollmächer,T., Auer,D.P., Functional MRI during sleep: BOLD signal decreases and

References 126

their electrophysiological correlates. European Journal of Neuroscience 20 (2004)

566-574.

[10] Kaufmann,C., Wehrle,R., Wetter,T.C., Holsboer,F., Auer,D.P.,

Pollmächer,T., Czisch,M., Brain activation and hypothalamic functional connectivity

during human non-rapid eye movement sleep: an EEG/fMRI study. Brain 129 (2006)

655-667.

[11] Laufs,H., Holt,J.L., Elfont,R., Krams,M., Paul,J.S., Krakow,K.,

Kleinschmidt,A., Where the BOLD signal goes when alpha EEG leaves. Neuroimage

31 (2006) 1408-1418.

[12] Bloch,F., Hansen,W.W., Packard,M., Nuclear Induction. Physical

Review 69 (1946) 127.

[13] Purcell,E.M., Torrey,H.C., Pound,R.V., Resonance Absorption by

Nuclear Magnetic Moments in A Solid. Physical Review 69 (1946) 37-38.

[14] Lauterbur,P.C., Image Formation by Induced Local Interactions -

Examples Employing Nuclear Magnetic-Resonance. Nature 242 (1973) 190-191.

[15] Haacke,E.M., Brown,R.W., Thompson,M.R. and Venkatesan,R.,

Magnetic resonance imaging: physical principles and sequence design, J. Wiley &

Sons, New York, (1999).

[16] Ogawa,S., Lee,T.M., Magnetic resonance imaging of blood vessels at

high fields: in vivo and in vitro measurements and image simulation. Magnetic

Resonance in Medicine 16 (1990) 9-18.

[17] Ogawa,S., Lee,T.M., Nayak,A.S., Glynn,P., Oxygenation-sensitive

contrast in magnetic resonance image of rodent brain at high magnetic fields.

Magnetic Resonance in Medicine 14 (1990) 68-78.

References 127

[18] Herson R., Büchel C., Josephs O. and Friston K., The Slice-Timing

Problem in Event-related fMRI. In: Human Brain Mapping Conference, (1999).

[19] van Strien N., MRI-tutorial: A neuroimaging repository (2009), Retrieved

from

http://www.mri-tutorial.com/index.php.

[20] MRIcro tutorial (2011), Retrieved from

http://www.cabiatl.com/mricro/mricro/mritut.html#ViewingIMG.

[21] Hutton, C, fMRI analysis: preprocessing: coresgistration and spatial

normalization (2-27-2009), Retrieved from

http://www.fil.ion.ucl.ac.uk/spm/doc/mfd/2009/.

[22] Brodmann,K., Vergleichende Lokalisationslehre der Grosshirnrinde in

ihren Prinzipien dargestellt auf Grund des Zellenbaues, Johann Ambrosius Barth

Verlag, Leipzig, (1909).

[23] Brodman areas (8-7-2010), Retrieved from

http://www.mrc-cbu.cam.ac.uk/people/jessica.grahn/neuroanatomy.html.

[24] Friston,K.J., Penny,W., Phillips,C., Kiebel,S., Hinton,G., Ashburner,J.,

Classical and Bayesian inference in neuroimaging: Theory. Neuroimage 16 (2002)

465-483.

[25] Friston,K., Holmes,A., Worsley,K., Poline,J., Frith,C., Frackowiak,R.,

Statistical parametric maps in functional imaging: A general linear approach. Human

Brain Mapping 2 (1994) 189-210.

[26] Raichle,M.E., A Paradigm Shift in Functional Brain Imaging. Journal of

Neuroscience 29 (2009) 12729-12734.

References 128

[27] McKeown,M.J., Makeig,S., Brown,G.G., Jung,T.P., Kindermann,S.S.,

Bell,A.J., Sejnowski,T.J., Analysis of fMRI data by blind separation into independent

spatial components. Human Brain Mapping 6 (1998) 160-188.

[28] Reijneveld,J.C., Ponten,S.C., Berendse,H.W., Stam,C.J., The

application of graph theoretical analysis to complex networks in the brain. Clinical

Neurophysiology 118 (2007) 2317-2331.

[29] Bullmore,E., Sporns,O., Complex brain networks: graph theoretical

analysis of structural and functional systems. Nature Reviews Neuroscience 10

(2009) 186-198.

[30] Berger,H., Über das Elektrenkephalogramm des Menschen. Archiv fur

Psychiatrie und Nervenkrankheiten 87 (1929) 527-570.

[31] Holmes,G. and Khazipov,R., Basic Neurophysiology and the Cortical

Basis of EEG. In: A.S.Blum, S.B.Rutkove (Eds.), The Clinical Neurophysiology

Primer, Humana Press, (2007) 19-33.

[32] da Silva,F.L., EEG: Origin and Measurement. In: C.Mulert, L.Lemieux

(Eds.), EEG-fMRI Physiological Basis, Technique, and Application, Springer, (2010)

20-35.

[33] Michel,C.M., Murray,M.M., Lantz,G.r., Gonzalez,S., Spinelli,L., Grave

de Peralta,R., EEG source imaging. Clinical Neurophysiology 115 (2004) 2195-2222.

[34] User Tutorial: EEG Measurement Setup (8-14-2009), Retrieved from

http://www.bci2000.org/wiki/index.php/User_Tutorial:EEG_Measurement_Setup.

[35] Easycap Products (1-25-2011), Retrieved from

http://www.easycap.de/easycap/e/products/products.htm.

References 129

[36] Nakamura,M., Chen,Q., Sugi,T., Ikeda,A., Shibasaki,H., Technical

quality evaluation of EEG recording based on electroencephalographers' knowledge.

Medical Engineering & Physics 27 (2005) 93-100.

[37] Hennighausen,E., Heil,M., Rosler,F., A Correction Method for Dc Drift

Artifacts. Electroencephalography and Clinical Neurophysiology 86 (1993) 199-204.

[38] Allen,P.J., EEG Instrumentation and Safety. In: C.Mulert, L.Lemieux

(Eds.), EEG-fMRI Physiological Basis, Technique, and Application, Springer, (2010)

115-133.

[39] Lemieux,L., Allen,P.J., Franconi,F., Symms,M.R., Fish,D.R., Recording

of EEG during fMRI experiments: Patient safety. Magnetic Resonance in Medicine 38

(1997) 943-952.

[40] Roth,B.J., Pascualleone,A., Cohen,L.G., Hallett,M., The Heating of

Metal-Electrodes During Rapid-Rate Magnetic Stimulation - A Possible Safety

Hazard. Electroencephalography and Clinical Neurophysiology 85 (1992) 116-123.

[41] Stevens,T.K., Ives,J.R., Klassen,L.M., Bartha,R., MR compatibility of

EEG scalp electrodes at 4 Tesla. Journal of Magnetic Resonance Imaging 25 (2007)

872-877.

[42] Mirsattari,S.M., Lee,D.H., Jones,D., Bihari,F., Ives,J.R., MRI compatible

EEG electrode system for routine use in the epilepsy monitoring unit and intensive

care unit. Clinical Neurophysiology 115 (2004) 2175-2180.

[43] Gutberlet, I, Did you know...? MR Correction, in Press Release (2007),

Retrieved from

http://www.brainproducts.com/pressrelease.php?tab=2&year=2007.

References 130

[44] Mandelkow,H., HaIder,P., Boesiger,P., Brandeis,D., Synchronization

facilitates removal of MRI artefacts from concurrent EEG recordings and increases

usable bandwidth. Neuroimage 32 (2006) 1120-1126.

[45] Allen,P.J., Josephs,O., Turner,R., A method for removing imaging

artifact from continuous EEG recorded during functional MRI. Neuroimage 12 (2000)

230-239.

[46] Ives,J.R., Warach,S., Schmitt,F., Edelman,R.R., Schomer,D.L.,

Monitoring the Patients Eeg During Echo-Planar Mri. Electroencephalography and

Clinical Neurophysiology 87 (1993) 417-420.

[47] Allen,P.J., Polizzi,G., Krakow,K., Fish,D.R., Lemieux,L., Identification of

EEG events in the MR scanner: The problem of pulse artifact and a method for its

subtraction. Neuroimage 8 (1998) 229-239.

[48] Yan,W.X., Mullinger,K.J., Geirsdottir,G.B., Bowtell,R., Physical

Modeling of Pulse Artefact Sources in Simultaneous EEG/fMRI. Human Brain

Mapping 31 (2010) 604-620.

[49] Briselli,E., Garreffa,G., Bianchi,L., Bianciardi,M., Macaluso,E.,

Abbafati,M., Grazia Marciani,M., Maraviglia,B., An independent component analysis-

based approach on ballistocardiogram artifact removing. Magnetic Resonance

Imaging 24 (2006) 393-400.

[50] Czisch,M., Wehrle,R., Stiegler,A., Peters,H., Andrade,K., Holsboer,F.,

Samann,P.G., Acoustic Oddball during NREM Sleep: A Combined EEG/fMRI Study.

Plos One 4 (2009) e6749.

[51] Cardoso,J.F., Blind signal separation: Statistical principles. Proceedings

of the IEEE 86 (1998) 2009-2025.

References 131

[52] Dang-Vu,T.T., Desseilles,M., Petit,D., Mazza,S., Montplaisir,J.,

Maquet,P., Neuroimaging in sleep medicine. Sleep Medicine 8 (2007) 349-372.

[53] Staner,L., Comorbidity of insomnia and depression. Sleep Medicine

Reviews 14 (2010) 35-46.

[54] Cohrs,S., Sleep Disturbances in Patients with Schizophrenia Impact

and Effect of Antipsychotics. CNS Drugs 22 (2008) 939-962.

[55] Briones,B., Adams,N., Strauss,M., Rosenberg,C., Whalen,C.,

Carskadon,M., Roebuck,T., Winters,M., Redline,S., Relationship between sleepiness

and general health status. Sleep 19 (1996) 583-588.

[56] Loomis,A.L., Harvey,E.N., Hobart,G.A., Cerebral states during sleep, as

studied by human brain potentials. Journal of Experimental Psychology 21 (1937)

127-144.

[57] Rechtschaffen,A., Kales,A.A., A manual of standardized terminology,

techniques and scoring system for sleep stages of human subjects. Washington D.

C. , NIH publication no. 204 (1968).

[58] Oken,B.S., Salinsky,M.C., Elsas,S.M., Vigilance, alertness, or sustained

attention: physiological basis and measurement. Clinical Neurophysiology 117 (2006)

1885-1901.

[59] Susmakova,K., Human Sleep and Sleep EEG. Measurement science

review 4 (2004) 59-74.

[60] Yang,C.M., Han,H.Y., Yang,M.H., Su,W.C., Lane,T., What subjective

experiences determine the perception of falling asleep during sleep onset period?

Consciousness and Cognition 19 (2010) 1084-1092.

References 132

[61] Ogilvie,R.D., Simons,I.A., Kuderian,R.H., Macdonald,T., Rustenburg,J.,

Behavioral, Event-Related Potential, and EEG FFT Changes at Sleep Onset.

Psychophysiology 28 (1991) 54-64.

[62] Jan,J.E., Reiter,R.J., Wasdell,M.B., Bax,M., The role of the thalamus in

sleep, pineal melatonin production, and circadian rhythm sleep disorders. Journal of

Pineal Research 46 (2009) 1-7.

[63] Czisch,M. and Wehrle,R., Sleep. In: C.Mulert, L.Lemieux (Eds.), EEG-

fMRI Physiological Basis, Technique, and Application, Springer, (2010) 279-305.

[64] Schabus,M., Dang-Vu,T.T., Albouy,G., Balteau,E., Boly,M., Carrier,J.,

Darsaud,A., Degueldre,C., Desseilles,M., Gais,S., Phillips,C., Rauchs,G.,

Schnakers,C., Sterpenich,V., Vandewalle,G., Luxen,A., Maquet,P., Hemodynamic

cerebral correlates of sleep spindles during human non-rapid eye movement sleep.

Proceedings of the National Academy of Sciences of the United States of America

104 (2007) 13164-13169.

[65] Cote,K.A., Epps,T.M., Campbell,K.B., The role of the spindle in human

information processing of high-intensity stimuli during sleep. Journal of Sleep

Research 9 (2000) 19-26.

[66] Steriade,M., McCormick,D.A., Sejnowski,T.J., Thalamocortical

oscillations in the sleeping and aroused brain. Science 262 (1993) 679-685.

[67] Bryant,P.A., Trinder,J., Curtis,N., Sick and tired: Does sleep have a vital

role in the immune system? Nature Reviews Immunology 4 (2004) 457-467.

[68] Maquet,P., Sleep function(s) and cerebral metabolism. Behavioural

Brain Research 69 (2007) 75-83.

[69] Vassalli,A., Dijk,D.J., Sleep function: current questions and new

approaches. European Journal of Neuroscience 29 (2009) 1830-1841.

References 133

[70] Axmacher,N., Draguhn,A., Elger,C.E., Fell,J., Memory processes during

sleep: beyond the standard consolidation theory. Cellular and Molecular Life

Sciences 66 (2009) 2285-2297.

[71] Born,J., Rasch,B., Gais,S., Sleep to remember. Neuroscientist 12

(2006) 410-424.

[72] Walker,M.P., The role of sleep in cognition and emotion. Annals of the

New York Academy of Sciences 1156 (2009) 168-197.

[73] Rasch,B., Buchel,C., Gais,S., Born,J., Odor cues during slow-wave

sleep prompt declarative memory consolidation. Science 315 (2007) 1426-1429.

[74] Sutherland,G.R., McNaughton,B., Memory trace reactivation in

hippocampal and neocortical neuronal ensembles. Current Opinion in Neurobiology

10 (2000) 180-186.

[75] Diekelmann,S., Born,J., The memory function of sleep. Nature Reviews

Neuroscience 11 (2010) 114-126.

[76] Walker,M.P., Stickgold,R., Sleep-dependent learning and memory

consolidation. Neuron 44 (2004) 121-133.

[77] Wagner,U., Gais,S., Born,J., Emotional memory formation is enhanced

across sleep intervals with high amounts of rapid eye movement sleep. Learning &

Memory 8 (2001) 112-119.

[78] Plihal,W., Born,J., Effects of early and late nocturnal sleep on priming

and spatial memory. Psychophysiology 36 (1999) 571-582.

[79] Plihal,W., Born,J., Effects of early and late nocturnal sleep on

declarative and procedural memory. Journal of Cognitive Neuroscience 9 (1997) 534-

547.

References 134

[80] Fogel,S.M., Smith,C.T., Cote,K.A., Dissociable learning-dependent

changes in REM and non-REM sleep in declarative and procedural memory systems.

Behavioural Brain Research 180 (2007) 48-61.

[81] Rauchs,G., Bertran,F., Guillery-Girard,B., Desgranges,B.,

Kerrouche,N., Denise,P., Foret,J., Eustache,F., Consolidation of strictly episodic

memories mainly requires rapid eye movement sleep. Sleep 27 (2004) 395-401.

[82] Gais,S., Plihal,W., Wagner,U., Born,J., Early sleep triggers memory for

early visual discrimination skills. Nature Neuroscience 3 (2000) 1335-1339.

[83] Rasch,B., Pommer,J., Diekelmann,S., Born,J., Pharmacological REM

sleep suppression paradoxically improves rather than impairs skill memory. Nature

Neuroscience 12 (2009) 396-397.

[84] Smith,C., Macneill,C., Impaired Motor Memory for A Pursuit Rotor Task

Following Stage-2 Sleep Loss in College-Students. Journal of Sleep Research 3

(1994) 206-213.

[85] Molle,M., Marshall,L., Gais,S., Born,J., Learning increases human

electroencephalographic coherence during subsequent slow sleep oscillations.


101 (2004) 13963-13968.

[86] Schabus,M., Gruber,G., Parapatics,S., Sauter,C., Klosch,G.,

Anderer,P., Klimesch,W., Saletu,B., Zeitlhofer,J., Sleep spindles and their

significance for declarative memory consolidation. Sleep 27 (2004) 1479-1485.

[87] Molle,M., Born,J., Hippocampus Whispering in Deep Sleep to Prefrontal

Cortex - For Good Memories? Neuron 61 (2009) 496-498.

[88] Durrant,S., Lewis,P.A., Memory Consolidation: Tracking Transfer with

Functional Connectivity. Current Biology 19 (2009) R860-R862.

References 135

[89] Girardeau,G., Benchenane,K., Wiener,S.I., Buzsaki,G., Zugaro,M.B.,

Selective suppression of hippocampal ripples impairs spatial memory. Nature

Neuroscience 12 (2009) 1222-1223.

[90] Hasselmo,M.E., McClelland,J.L., Neural models of memory. Current

Opinion in Neurobiology 9 (1999) 184-188.

[91] Squire,L.R., Memory systems of the brain: a brief history and current

perspective. Neurobiology of Learning and Memory 82 (2004) 171-177.

[92] Takashima,A., Nieuwenhuis,I.L., Jensen,O., Talamini,L.M.,

Rijpkema,M., Fernandez,G., Shift from hippocampal to neocortical centered retrieval

network with consolidation. Journal of Neuroscience 29 (2009) 10087-10093.

[93] Clemens,Z., Molle,M., Eross,L., Barsi,P., Halasz,P., Born,J., Temporal

coupling of parahippocampal ripples, sleep spindles and slow oscillations in humans.

Brain 130 (2007) 2868-2878.

[94] Siapas,A.G., Wilson,M.A., Coordinated interactions between

hippocampal ripples and cortical spindles during slow-wave sleep. Neuron 21 (1998)

1123-1128.

[95] O'Neill,J., Pleydell-Bouverie,B., Dupret,D., Csicsvari,J., Play it again:

reactivation of waking experience and memory. Trends in Neurosciences 33 (2010)

220-229.

[96] Buckner,R.L., The role of the hippocampus in prediction and

imagination. Annual Review of Psychology 61 (2010) 27-28.

[97] Amunts,K., Kedo,O., Kindler,M., Pieperhoff,P., Mohlberg,H., Shah,N.J.,

Habel,U., Schneider,F., Zilles,K., Cytoarchitectonic mapping of the human amygdala,

hippocampal region and entorhinal cortex: intersubject variability and probability

maps. Anatomy and Embryology (Berl) 210 (2005) 343-352.

References 136

[98] McNaughton,N., The role of the subiculum within the behavioural

inhibition system. Behavioural Brain Research 174 (2006) 232-250.

[99] Levita,L., Muzzio,I.A., Role of the hippocampus in goal-oriented tasks

requiring retrieval of spatial versus non-spatial information. Neurobiology of Learning

and Memory 93 (2010) 581-588.

[100] Tejada,S., Gonzalez,J.J., Rial,R.V., Coenen,A.M.L., Gamundi,A.,

Esteban,S., Electroencephalogram Functional Connectivity Between Rat

Hippocampus and Cortex After Pilocarpine Treatment. Neuroscience 165 (2010) 621-

631.

[101] Witter,M.P., Connectivity of the Hippocampus. In: V.Cutsuridis,

B.Graham, S.Cobb, I.Vida (Eds.), Hippocampal Microcircuits, Springer, New York,

(2010) 5-26.

[102] Fox,M.D., Raichle,M.E., Spontaneous fluctuations in brain activity

observed with functional magnetic resonance imaging. Nature Reviews

Neurosciscience 8 (2007) 700-711.

[103] Jann,K., Kottlow,M., Dierks,T., Boesch,C., Koenig,T., Topographic

Electrophysiological Signatures of fMRI Resting State Networks. Plos One 5 (2010)

e12945.

[104] Buckner,R.L., Andrews-Hanna,J.R., Schacter,D.L., The brain's default

network: anatomy, function, and relevance to disease. Annals of the New York

Academy of Sciences 1124 (2008) 1-38.

[105] Vincent,J.L., Snyder,A.Z., Fox,M.D., Shannon,B.J., Andrews,J.R.,

Raichle,M.E., Buckner,R.L., Coherent spontaneous activity identifies a hippocampal-

parietal memory network. Journal of Neurophysiology 96 (2006) 3517-3531.

References 137

[106] Larson-Prior,L.J., Zempel,J.M., Nolan,T.S., Prior,F.W., Snyder,A.Z.,

Raichle,M.E., Cortical network functional connectivity in the descent to sleep.


106 (2009) 4489-4494.

[107] Horovitz,S.G., Braun,A.R., Carr,W.S., Picchioni,D., Balkin,T.J.,

Fukunaga,M., Duyn,J.H., Decoupling of the brain's default mode network during deep

sleep. Proceedings of the National Academy of Sciences of the United States of

America 106 (2009) 11376-11381.

[108] Sämann,P.G., Wehrle,R., Spoormaker,V., Hoehn,D., Peters,H.,

Holsboer,F., Czisch,M., Development of the brain default mode network from

wakefulness into slow wave sleep. Cerebral Cortex (2011) in press.

[109] Eickhoff,S.B., Stephan,K.E., Mohlberg,H., Grefkes,C., Fink,G.R.,

Amunts,K., Zilles,K., A new SPM toolbox for combining probabilistic cytoarchitectonic

maps and functional imaging data. Neuroimage 25 (2005) 1325-1335.

[110] Brett,M., Anton,J.-L., Valabregue,R., Poline,J.-P., Region of interest

analysis using an SPM toolbox. Presented at the 8th International Conference on

Functional Mapping of the Human Brain, Sendai, Japan (2002).

[111] Hayasaka,S., Phan,K.L., Liberzon,I., Worsley,K.J., Nichols,T.E.,

Nonstationary cluster-size inference with random field and permutation methods.

Neuroimage 22 (2004) 676-687.

[112] De Gennaro,L., Ferrara,M., Sleep spindles: an overview. Sleep

Medicine Reviews 7 (2003) 423-440.

[113] Friston,K.J., Stephan,K.E., Lund,T.E., Morcom,A., Kiebel,S., Mixed-

effects and fMRI studies. Neuroimage 24 (2005) 244-252.

References 138

[114] Spoormaker,V.I., Schroter,M.S., Gleiser,P.M., Andrade,K.C.,

Dresler,M., Wehrle,R., Sämann,P.G., Czisch,M., Development of a Large-Scale

Functional Brain Network during Human Non-Rapid Eye Movement Sleep. Journal of

Neuroscience 30 (2010) 11379-11387.

[115] Svoboda,E., McKinnon,M.C., Levine,B., The functional neuroanatomy

of autobiographical memory: a meta-analysis. Neuropsychologia 44 (2006) 2189-

2208.

[116] Wagner,A.D., Shannon,B.J., Kahn,I., Buckner,R.L., Parietal lobe

contributions to episodic memory retrieval. Trends in Cognitive Sciences 9 (2005)

445-453.

[117] Burianova,H., McIntosh,A.R., Grady,C.L., A common functional brain

network for autobiographical, episodic, and semantic memory retrieval. Neuroimage.

49 (2010) 865-874.

[118] Andrews-Hanna,J.R., Reidler,J.S., Sepulcre,J., Poulin,R., Buckner,R.L.,

Functional-anatomic fractionation of the brain's default network. Neuron 65 (2010)

550-562.

[119] Martuzzi,R., Ramani,R., Qiu,M., Rajeevan,N., Constable,R.T.,

Functional connectivity and alterations in baseline brain state in humans.

Neuroimage. 49 (2010) 823-834.

[120] Gais,S., Born,J., Low acetylcholine during slow-wave sleep is critical for

declarative memory consolidation. Proceedings of the National Academy of Sciences

of the United States of America 101 (2004) 2140-2144.

[121] Peigneux,P., Laureys,S., Fuchs,S., Collette,F., Perrin,F., Reggers,J.,

Phillips,C., Degueldre,C., Del Fiore,G., Aerts,J., Luxen,A., Maquet,P., Are spatial

References 139

memories strengthened in the human hippocampus during slow wave sleep? Neuron

44 (2004) 535-545.

[122] Genzel,L., Dresler,M., Wehrle,R., Grozinger,M., Steiger,A., Slow wave

sleep and REM sleep awakenings do not affect sleep dependent memory

consolidation. Sleep 32 (2009) 302-310.

[123] Gais,S., Molle,M., Helms,K., Born,J., Learning-dependent increases in

sleep spindle density. Journal of Neuroscience 22 (2002) 6830-6834.

[124] Nishida,M., Walker,M.P., Daytime naps, motor memory consolidation

and regionally specific sleep spindles. Plos One 2 (2007) e341.

[125] Ribeiro,S., Gervasoni,D., Soares,E.S., Zhou,Y., Lin,S.C., Pantoja,J.,

Lavine,M., Nicolelis,M.A., Long-lasting novelty-induced neuronal reverberation during

slow-wave sleep in multiple forebrain areas. PLoS Biology 2 (2004) e24.

[126] Payne,J.D., Kensinger,E.A., Sleep Leads to Changes in the Emotional

Memory Trace: Evidence from fMRI. Journal of Cognitive Neuroscience (2010) epub

ahead of print.

[127] Darsaud,A., Wagner,U., Balteau,E., Desseilles,M., Sterpenich,V.,

Vandewalle,G., Albouy,G., Dang-Vu,T., Collette,F., Boly,M., Schabus,M.,

Degueldre,C., Luxen,A., Maquet,P., Neural Precursors of Delayed Insight. Journal of

Cognitive Neuroscience (2010) epub ahead of print.

[128] Darsaud,A., Dehon,H., Lahl,O., Sterpenich,V., Boly,M., Dang-Vu,T.,

Desseilles,M., Gais,S., Matarazzo,L., Peters,F., Schabus,M., Schmidt,C.,

Tinguely,G., Vandewalle,G., Luxen,A., Maquet,P., Collette,F., Does Sleep Promote

False Memories? Journal of Cognitive Neuroscience 23 (2011) 26-40.

[129] Sterpenich,V., Albouy,G., Darsaud,A., Schmidt,C., Vandewalle,G.,

Dang Vu,T.T., Desseilles,M., Phillips,C., Degueldre,C., Balteau,E., Collette,F.,

References 140

Luxen,A., Maquet,P., Sleep promotes the neural reorganization of remote emotional

memory. Journal of Neuroscience 29 (2009) 5143-5152.

[130] Rauchs,G., Orban,P., Schmidt,C., Albouy,G., Balteau,E., Degueldre,C.,

Schnackers,C., Sterpenich,V., Tinguely,G., Luxen,A., Maquet,P., Peigneux,P., Sleep

modulates the neural substrates of both spatial and contextual memory consolidation.

Plos One 3 (2008) e2949.

[131] Gais,S., Albouy,G., Boly,M., ng-Vu,T.T., Darsaud,A., Desseilles,M.,

Rauchs,G., Schabus,M., Sterpenich,V., Vandewalle,G., Maquet,P., Peigneux,P.,

Sleep transforms the cerebral trace of declarative memories. Proceedings of the

National Academy of Sciences of the United States of America 104 (2007) 18778-

18783.

[132] Wagner,T., Axmacher,N., Lehnertz,K., Elger,C.E., Fell,J., Sleep-

dependent directional coupling between human neocortex and hippocampus. Cortex

46 (2010) 256-263.

[133] Sirota,A., Csicsvari,J., Buhl,D., Buzsaki,G., Communication between

neocortex and hippocampus during sleep in rodents. Proceedings of the National

Academy of Sciences of the United States of America 100 (2003) 2065-2069.

[134] Battaglia,F.P., Sutherland,G.R., McNaughton,B.L., Hippocampal sharp

wave bursts coincide with neocortical "up-state" transitions. Learning & Memory 11

(2004) 697-704.

[135] Axmacher,N., Mormann,F., Fernandez,G., Elger,C.E., Fell,J., Memory

formation by neuronal synchronization. Brain Research Reviews 52 (2006) 170-182.

[136] Schabus,M., Hoedlmoser,K., Pecherstorfer,T., Anderer,P., Gruber,G.,

Parapatics,S., Sauter,C., Kloesch,G., Klimesch,W., Saletu,B., Zeitlhofer,J.,

References 141

Interindividual sleep spindle differences and their relation to learning-related

enhancements. Brain Research 1191 (2008) 127-135.

[137] Fogel,S.M., Smith,C.T., Learning-dependent changes in sleep spindles

and Stage 2 sleep. Journal of Sleep Research 15 (2006) 250-255.

[138] Clemens,Z., Fabo,D., Halasz,P., Overnight verbal memory retention

correlates with the number of sleep spindles. Neuroscience 132 (2005) 529-535.

[139] Herkenham,M., The connections of the nucleus reuniens thalami:

evidence for a direct thalamo-hippocampal pathway in the rat. Journal of

Comparative Neurology 177 (1978) 589-610.

[140] Wouterlood,F.G., Saldana,E., Witter,M.P., Projection from the nucleus

reuniens thalami to the hippocampal region: light and electron microscopic tracing

study in the rat with the anterograde tracer Phaseolus vulgaris-leucoagglutinin.

Journal of Comparative Neurology 296 (1990) 179-203.

[141] Tononi,G., Cirelli,C., Sleep and synaptic homeostasis: a hypothesis.

Brain Research Bulletin 62 (2003) 143-150.

[142] Dalbokova,D., Gille,H.G., Ullsperger,P., Amplitude Variations in P300

Component Due to Unpredictable Stepwise Change of Stimulus Probability.

International Journal of Psychophysiology 10 (1990) 33-38.

[143] Polich,J., Alexander,J.E., Bauer,L.O., Kuperman,S., Morzorati,S.,

OConnor,S.J., Porjesz,B., Rohrbaugh,J., Begleiter,H., P300 topography of

amplitude/latency correlations. Brain Topography 9 (1997) 275-282.

[144] Ma,Q.G., Wang,X.Y., Shu,L.C., Dai,S.Y., P300 and categorization in

brand extension. Neuroscience Letters 431 (2008) 57-61.

[145] Polich,J., Updating p300: An integrative theory of P3a and P3b. Clinical


References 142

[146] Rosburg,T., Boutros,N.N., Ford,J.M., Reduced auditory evoked

potential component N100 in schizophrenia - A critical review. Psychiatry Research

161 (2008) 259-274.

[147] Mulert,C., Seifert,C., Leicht,G., Kirsch,V., Ertl,M., Karch,S.,

Moosmann,M., Lutz,J., Moller,H.J., Hegerl,U., Pogarell,O., Jager,L., Single-trial

coupling of EEG and fMRI reveals the involvement of early anterior cingulate cortex

activation in effortful decision making. Neuroimage 42 (2008) 158-168.

[148] Crowley,K.E., Colrain,I.M., A review of the evidence for P2 being an

independent component process: age, sleep and modality. Clinical Neurophysiology

115 (2004) 732-744.

[149] Acharya,R., Faust,O., Kannathal,N., Chua,T., Laxminarayan,S., Non-

linear analysis of EEG signals at various sleep stages. Computer Methods and

Programs in Biomedicine 80 (2005) 37-45.

[150] Ruby,P., Caclin,A., Boulet,S., Delpuech,C., Morlet,D., Odd sound

processing in the sleeping brain. Journal of Cognitive Neuroscience 20 (2008) 296-

311.

[151] Mander,B.A., Reid,K.J., Davuluri,V.K., Small,D.M., Parrish,T.B.,

Mesulam,M.M., Zee,P.C., Gitelman,D.R., Sleep deprivation alters functioning within

the neural network underlying the covert orienting of attention. Brain Research 1217

(2008) 148-156.

[152] Benar,C.G., Schon,D., Grimault,S., Nazarian,B., Burle,B., Roth,M.,

Badier,J.M., Marquis,P., Liegeois-Chauvel,C., Anton,J.L., Single-trial analysis of

oddball event-related potentials in simultaneous EEG-fMRI. Human Brain Mapping

28 (2007) 602-613.

References 143

[153] Debener,S., Ullsperger,M., Siegel,M., Fiehler,K., von Cramon,D.Y.,

Engel,A.K., Trial-by-trial coupling of concurrent electroencephalogram and functional

magnetic resonance imaging identifies the dynamics of performance monitoring.

Journal of Neuroscience 25 (2005) 11730-11737.

[154] Iyer,D., Zouridakis,G., Single-trial evoked potential estimation:

Comparison between independent component analysis and wavelet denoising.

Clinical Neurophysiology 118 (2007) 495-504.

[155] Bagshaw,A.P., Warbrick,T., Single trial variability of EEG and fMRI

responses to visual stimuli. Neuroimage 38 (2007) 280-292.

[156] Zouridakis,G., Iyer,D., Diaz,J., Patidar,U., Estimation of individual

evoked potential components using iterative independent component analysis.

Physics in Medicine and Biology 52 (2007) 5353-5368.

[157] Mantini,D., Corbetta,M., Perrucci,M.G., Romani,G.L., Del Gratta,C.,

Large-scale brain networks account for sustained and transient activity during target

detection. Neuroimage 44 (2009) 265-274.

[158] Mutihac,R., Mutihac,R.C., A comparative study of independent

component analysis algorithms for electroencephalography. Romanian Reports in

Physics 59 (2007) 831-860.

[159] Zbilut,J.P., Webber,C.L., Embeddings and Delays As Derived from

Quantification of Recurrence Plots. Physics Letters A 171 (1992) 199-203.

[160] Marwan,N., Encounters With Neighbours-Current Development of

Concepts Based on Recurrence Plots And Their Application. In: Fakuktät

Mathematik Und Naturwissenschaften Universität Potsdam, (2003).

References 144

[161] Carrubba,S., Frilot,C., Chesson,A.L., Marino,A.A., Nonlinear EEG

activation evoked by low-strength low-frequency magnetic fields. Neuroscience

Letters 417 (2007) 212-216.

[162] Carrubba,S., Frilot,C., Chesson,A.L., Marino,A.A., Evidence of a

nonlinear human magnetic sense. Neuroscience 144 (2007) 356-367.

[163] Carrubba,S., Frilot,C., Chesson,A.L., Webber,C.L., Zbilut,J.P.,

Marino,A.A., Magnetosensory evoked potentials: Consistent nonlinear phenomena.

Neuroscience Research 60 (2008) 95-105.

[164] Schinkel,S., Marwan,N., Kurths,J., Brain signal analysis based on

recurrences. Journal of Physiology (Paris) 103 (2009) 315-323.

[165] Marwan,N., Meinke,A., Extended recurrence plot analysis and its

application to ERP data. International Journal of Bifurcation and Chaos 14 (2004)

761-771.

[166] Zenker,F., Barajas,J.J., Auditory P300 development from an active,

passive and single-tone paradigms. International Journal of Psychophysiology 33

(1999) 99-111.

[167] Lindin,M., Zurron,M., Diaz,F., Changes in P300 amplitude during an

active standard auditory oddball task. Biological Psychology 66 (2004) 153-167.

[168] Barry,R.J., Rushby,J.A., Smith,J.L., Clarke,A.R., Croft,R.J.,

Wallace,M.J., Brain dynamics in the active vs. passive auditory oddball task:

Exploration of narrow-band EEG phase effects. Clinical Neurophysiology 118 (2007)

2234-2247.

[169] Polich,J., Attention, Probability, and Task Demands As Determinants of

P300 Latency from Auditory-Stimuli. Electroencephalography and Clinical


References 145

[170] Stevens,M.C., Laurens,K.R., Liddle,P.F., Kiehl,K.A., The

hemodynamics of oddball processing during single-tone and two-tone target

detection tasks. International Journal of Psychophysiology 60 (2006) 292-303.

[171] Takens,F., Detecting Strange Attractors in Turbulence. In: D.A.Rand,

L.S.Young (Eds.), Lectures Notes in Mathematics, Vol. 898, Springer, Berlin, (1981)

366-381.

[172] Eckmann,J.P., Oliffson Kamphorst,S. and Ruelle,D., Recurrence Plots

of Dynamical Systems. In: (1987), 973-977.

[173] Marwan,N., Romano,M.C., Thiel,M., Kurths,J., Recurrence plots for the

analysis of complex systems. Physics Reports-Review Section of Physics Letters 438

(2007) 237-329.

[174] Marwan,N., Wessel,N., Meyerfeldt,U., Schirdewan,A., Kurths,J.,

Recurrence-plot-based measures of complexity and their application to heart-rate-

variability data. Physical Review e 66 (2002) e026702.

[175] Kennel,M.B., Brown,R., Abarbanel,H.D.I., Determining Embedding

Dimension for Phase-Space Reconstruction Using A Geometrical Construction.

Physical Review A 45 (1992) 3403-3411.

[176] Fraser,A.M., Swinney,H.L., Independent Coordinates for Strange

Attractors from Mutual Information. Physical Review A 33 (1986) 1134-1140.

[177] Demiralp,T., Ademoglu,A., Schurmann,M., Basar-Eroglu,C., Basar,E.,

Detection of P300 waves in single trials by the wavelet transform (WT). Brain and

Language 66 (1999) 108-128.

[178] Schinkel,S., Marwan,N., Kurths,J., Order patterns recurrence plots in

the analysis of ERP data. Cognitive Neurodynamics 1 (2007) 317-325.

References 146

[179] Marwan,N., Groth,A., Kurths,J., Quantification of Order Patterns

Recurrence Plots of Event Related Potentials. Chaos and Complexity Letters 2

(2007) 301-314.

[180] Bennington,J.Y., Polich,J., Comparison of P300 from passive and

active tasks for auditory and visual stimuli. International Journal of Psychophysiology

34 (1999) 171-177.

[181] Horovitz,S.G., Skudlarski,P., Gore,J.C., Correlations and dissociations

between BOLD signal and P300 amplitude in an auditory oddball task: a parametric

approach to combining fMRI and ERP. Magnetic Resonance Imaging 20 (2002) 319-

325.

[182] Hirata,K., Lehmann,D., N1 and P2 of Frequent and Rate Event-Related

Potentials Show Effects and After-Effects of the Attended Target in the Oddball-

Paradigm. International Journal of Psychophysiology 9 (1990) 293-301.

[183] Starr,A., Aguinaldo,T., Roe,M., Michalewski,H.J., Sequential changes of

auditory processing during target detection: Motor responding versus mental

counting. Electromyography and Motor Control-Electroencephalography and Clinical


[184] Iwanski,J.S., Bradley,E., Recurrence plots of experimental data: To

embed or not to embed? Chaos 8 (1998) 861-871.

[185] Eguiluz,V.M., Chialvo,D.R., Cecchi,G.A., Baliki,M., Apkarian,A.V.,

Scale-free brain functional networks. Physical Review Letters 94 (2005).

[186] Friston,K.J., The labile brain. I. Neuronal transients and nonlinear

coupling. Philosophical Transactions of the Royal Society B-Biological Sciences 355

(2000) 215-236.

References 147

[187] Rulkov,N.F., Sushchik,M.M., Tsimring,L.S., Abarbanel,H.D.I.,

Generalized Synchronization of Chaos in Directionally Coupled Chaotic Systems.

Physical Review e 51 (1995) 980-994.

[188] Romano,M.C., Thiel,M., Kurths,J., von Bloh,W., Multivariate recurrence

plots. Physics Letters A 330 (2004) 214-223.

148

ACKNOWLEDGEMENTS

I thank Prof. Stephan Paul for accepting me as a PhD student, for being kind,

open and interested in my work.

I thank also Prof. J. Leo van Hemmen and Prof. Sibylle Ziegler, for being part

of my examiner committee.

I consider myself lucky to always find the best people to work with/for, and

this applies to the entire neuroimaging group.

I would like to particularly thank Michael Czisch, for being a superb advisor

and knowing when to criticize and when to push me, for giving me the space and

time to walk on my own feet, for supporting me and sometimes even believing in me

more than myself. I am extremely grateful and I will always be.

Philipp Sämann for always knowing how the stretch my limits, for everything I

learned and for giving your best to improve any work on the group.

Renate Wehrle, I thank you not just for all scientific advice and what I have

learned from you but also for being so kind and friendly since the first day I arrived in

Germany, for the care and help that I needed to feel more at home.

Victor Spoormaker, you are particularly gifted and add a lot to the group. I

learned a lot with you and I thank you for that. Thank you also for your friendship,

availability and support.

For the friendship, help, advice and making my stay here much more

pleasant, I thank Sara Kiem, Roberto Goya-Maldonado and Barbara Grünecker.

Acknowledgements 149

For the “animals”, “zombies fans” and “memory gifted”, Sebastian Kaltwasser,

Benedikt Bedenk, Martin Dresler and Boris-Nikolai Konrad, I thank you for making life

even more interesting.

It has been also a pleasure to work with competent and welcoming

technicians and technical assistants Rosa Schirmer, Reinhold Borschke, Elke

Schreiter, Armin Mann, Stephanie Alam and Ines Eidner. Thank you for this

background support.

I thank the director of the MPI, Prof. Florian Holsboer, for the opportunity to

work in a so rich environment, with amazing researchers and structural facilities.

As important as the co-workers, my family was primordial in all my steps here

and will always be in any step I will still give in my life.

To my mom, Maria Auxiliadora, for being so perfect and taking such good

care of me, even from so far away. Thank you for calling me almost every day just to

say good night and making sure I was fine. Thank you for being unconditionally by

my side.

To my sisters, Fábia and Cybelle, who I miss a lot, thank you for cheering me

up and always being there for me. I cannot forget to thank Bartolomeu Silva for

making me laugh, for your friendship and being an excellent host.

To my father, J.M Andrade thank for the support and care.

I am especially grateful to my partner and friend, Danusa Colares, for

delaying and leaving many things of your life to be here with me. It was really hard

but you gave me strength and went through with me anyway. I love everything you

are and everything you make me be. Thank you for being part of my life since I was

twelve.

150

CONTRIBUTIONS

1. Recurrence quantification analysis allows for single-trial stimulus differentiation

in evoked potential studies.

Andrade KC, Wehrle R, Spoormaker, VI, SämannPG, Czisch M.

Clinical Neurophysiology (under revision).

2. Sleep spindles and hippocampal functional connectivity in human NREM

sleep.

Andrade KC, Spoormaker VI, Dresler M, Wehrle R, Holsboer F, Sämann PG, Czisch M.

The Journal of Neuroscience (under revision).

3. The neural correlates of negative prediction error signaling in human fear

conditioning.

Spoormaker VI, Andrade KC, Schröter MS, Sturm A, Goya-Maldonado R, Sämann PG,

Czisch M.

Neuroimage, 54(3):2250-6 (2011)

4. The neural correlates and temporal sequence of the relationship between

shock exposure, disturbed sleep and impaired consolidation of fear extinction.

Spoormaker VI, Sturm A, Andrade KC, Schröter MS, Goya-Maldonado R, Holsboer F, Wetter

TC, Sämann PG, Czisch M.

Journal of Psychiatric Research, 44(16):1121-8 (2010)

Contributions 151

5. Development of a Large-Scale Functional Brain Network during Human Non-

Rapid Eye Movement Sleep.

Spoormaker VI, Schroter MS, Gleiser PM, Andrade KC, Dresler M, Wehrle R, Sämann PG,

Czisch M.

The Journal of Neuroscience, 30(34):11379-87 (2010)

6. Acoustic oddball during NREM sleep: a combined EEG/fMRI study.

Czisch M, Wehrle R, Stiegler A, Peters H, Andrade KC, Holsboer F, Sämann PG.

PLoS One 4(8):e6749 (2009)

simultaneous functional magnetic resonance …mediatum.ub.tum.de/doc/1070878/932582.pdfsimultaneous...

Documents